Arxiv:Math/0702787V3 [Math.PR] 6 Oct 2007 Ote F E Cecse Ehius 6 Ot Ega.F- Gray

arXiv:math/0702787v3 [math.PR] 6 Oct 2007 ote F e cecse ehius 6 ot eGa.F- Gray. de route 16, Techniques. fcomte.fr et Sciences des Comté, UFR l eeaiaino oeo h eut n[8]i h follow the in modeli [B81] are in we results variat (i) that the the system of and the some processes of of space valued generalization phase real a the of in space values the with in image its var associated various and interest. action of o valued processes refined real subsequently a been formulate has to that and [Ne67] in introduced precise not variable. are random a parameters of th whose to realizations or adapted perturbations framework random a establishing to con is the goal within the mechanics instances quantum A by exhibited TZ97a, effects TZ97, mechanic random classical ZM84, of of generalization ZY82, pieces stochastic some suitable Y81, behind motivations B81, The therein). [Ne67, of references of instance context theory the for the Itô introduced K. (see to since mechanics ever subject classical research of generalization The Introduction 1 Keywords: [email protected] 2 1 h prahfloe nti ae scoe oteoeintro one the to closer is paper this in followed approach The pr of class a use category first the in work of pieces the of Most eteNtoa el ehrh cetfiu,Département Scientifique, Recherche la de National Centre eatmnod ´sc eoia nvria eZaragoza de Universidad Teórica. F´ısica de Departamento emk xesv s ftegoa tcatcaayi tool analysis stochastic global the of use extensive make We h ri pcsotie fe euto r eeial n generically param are of reduction ev is. also may after space is they obtained that it spaces symmetries orbit but the the to handled respect be with can them space reducing that phase systems non-Euclidean of handle to spectrum [Sch82] Schwartz L. and nonee ncasclmcais eea etrsade and presented. features are equations Several criti these natural of mechanics. semimartingales a classical by in equations characterized encountered Hamilton are the manifolds, of symplectic generalization stochastic a down euetegoa tcatcaayi ol nrdcdb P. by introduced tools analysis stochastic global the use We tcatcHmloindnmclsystems dynamical Hamiltonian Stochastic tcatcHmlo qain,sohsi aitoa pr variational stochastic equations, Hamilton stochastic onAde Lázaro-Cam´ıJoan-Andreu Abstract 1 1 hudpoiea xlnto fteintrinsically the of explanation an provide should s 53 ea¸o ee.France. Besan¸con cedex. 25030 n unPboOrtega Juan-Pablo and eMt´mtqe eBsncn nvri´ eFranche- Besan¸con, Université de Mathématiques de de er ebn,1.E509Zrgz.Spain. Zaragoza. E-50009 12. Cerbuna, Pedro . okrltdt hsfil a ntehp hta that hope the in lay field this to related work tcatcdffrnileutosi h 1950s the in equations differential stochastic adigo ehnclssessubjected systems mechanical of handling e n directions: ing ainlpicpe hs xrml r the are extremals whose principles iational a cinpicpesmlrt h one the to similar principle action cal e h er.Ti eiaiecnb used be can derivative This years. the ver ape nrlto ihtesolution the with relation in xamples tcatcdnmc a ena active an been has dynamics stochastic naPisnmnfl ht o exact for that, manifold Poisson a on ydtrie n r ec oee as modeled hence are and determined ly osaetkni h pc fprocesses of space the in taken are ions eto h hoyo iuin nother In . diffusions of theory the of text nrdcdb .A ee M1 M82] [M81, Meyer A. P. by introduced s nEcien vni h rgnlphase original the if even on-Euclidean, cse hthv tcatcderivative stochastic a have that ocesses g hspprcnb culyse as seen actually be can paper This ng. ue n[8]i hc h cinhas action the which in [B81] in duced .MyradL cwrzt write to Schwartz L. and Meyer A. .Ti etr o nywdn the widens only not feature This s. 3 D6 R0,BO7] and BRO07a], BRO07, CD06, 03, nulyhv se[O7) indeed, [LO07]); (see have entually nil,sohsi mechanics. stochastic inciple, utiprac ttetm of time the at importance ount J 2 uan-Pablo.Ortega@univ- Lázaro and Ortega: Stochastic Hamiltonian dynamical systems 2

(ii) The stochastic dynamical components of the system are modeled by continuous semimartingales and are not limited to Brownian motion. (iii) We handle stochastic Hamiltonian systems on Poisson manifolds and not only on symplectic manifolds. (iv) The variational principle that we propose in Theorem 4.14 is not just satisfied by the stochastic Hamiltonian equations (as in [B81]) but fully characterizes them. There are various reasons that have lead us to consider these generalized Hamiltonian systems. First, even though the laws that govern the dynamics of classical mechanical systems are, in principle, completely known, the finite precision of experimental measurements yields impossible the estimation of the parameters of a particular given one with total accuracy. Second, the modeling of complex physical systems involves most of the time simplifying assumptions or idealizations of parts of the system, some of which could be included in the description as a stochastic component; this modeling philosophy has been extremely successful in the social sciences [BJ76]. Third, even if the model and the parameters of the system are known with complete accuracy, the solutions of the associated differential equations may be of great complexity and exhibit high sensitivity to the initial conditions hence making the probabilistic treatment and description of the solutions appropriate. Finally, we will see (Section 3.3) how stochastic Hamiltonian modeling of microscopic systems can be used to model dissipation and macroscopic damping. The paper is structured as follows: in Section 2 we introduce the stochastic Hamilton equations with phase space a given Poisson manifold and we study some of the fundamental properties of the solution semimartingales like, for instance, the preservation of symplectic leaves or the characterization of the conserved quantities. This section contains a discussion on two notions on non-linear stability, almost sure Lyapunov stability and stability in probability, that reduce in the deterministic setup to the standard definition of Lyapunov stability. We formulate criteria that generalize to the Hamiltonian stochastic context the standard energy methods to conclude the stability of a Hamiltonian equilibrium using existing conservation laws. More specifically, there are two different natural notions of conserved quantity in the stochastic context that, via a stochastic Dirichlet criterion (Theorem 2.15) allow one to conclude the different kinds of stability that we have mentioned above. Section 3 contains several examples: in the first one we show how the systems studied by Bismut in [B81] fall in the category introduced in Section 2. We also see that a damped oscillator can be described as the average motion of the solution semimartingale of a natural stochastic Hamiltonian system, and that Brownian motion in a manifold is the projection onto the base space of very simple Hamiltonian stochastic semimartingale defined on the cotangent bundle of the manifold or of its orthonormal frame bundle, depending on the availability or not of a parallelization for the manifold in question. Section 4 is dedicated to showing that the stochastic Hamilton equations are characterized by a critical action principle that generalizes the one found in the treatment of deterministic systems. In order to make this part more readable, the proofs of most of the technical results needed to prove the theorems in this section have been included separately at the end of the paper. One of the goals of this paper is conveying to the geometric mechanics community the plentitude of global tools available to handle mechanical problems that contain a stochastic component and that do not seem to have been exploited to the full extent of their potential. In order to facilitate the task of understanding the paper to non-probabilists we have included an appendix that provides a self-contained presentation of some major facts in stochastic calculus on manifolds needed for a first comprehension of our results. Those pages are a very short and superficial presentation of a deep and technical field of mathematics so the reader interested in a more complete account is encouraged to check with the references quoted in the appendix and especially with the excellent monograph [E89]. Lázaro and Ortega: Stochastic Hamiltonian dynamical systems 3

Conventions: All the manifolds in this paper are ﬁnite dimensional, second-countable, locally compact, and Hausdorﬀ (and hence paracompact).

2 The stochastic Hamilton equations

In this section we present a natural generalization of the standard Hamilton equations in the stochastic context. Even though the arguments gathered in the following paragraphs as motivation for these equations are of formal nature, we will see later on that, as it was already the case for the standard Hamilton equations, they satisfy a natural variational principle. We recall that a symplectic manifold is a pair (M, ω), where M is a manifold and ω ∈ Ω2(M) is a closed non-degenerate two-form on M, that is, dω = 0 and, for every m ∈ M, the map v ∈ ∗ TmM 7→ ω(m)(v, ·) ∈ TmM is a linear isomorphism between the tangent space TmM to M at m and ∗ the cotangent space TmM. Using the nondegeneracy of the symplectic form ω, one can associate each ∞ function h ∈ C (M) a vector ﬁeld Xh ∈ X(M), deﬁned by the equality

iXh ω = dh. (2.1)

We will say that Xh is the Hamiltonian vector field associated to the Hamiltonian function h. The expression (2.1) is referred to as the Hamilton equations. A Poisson manifold is a pair (M, {·, ·}), where M is a manifold and {·, ·} is a bilinear operation on C∞(M) such that (C∞(M), {·, ·}) is a Lie algebra and {·, ·} is a derivation (that is, the Leibniz identity holds) in each argument. The functions in the center C(M) of the Lie algebra (C∞(M), {·, ·}) are called Casimir functions. From the natural isomorphism between derivations on C∞(M) and vector fields ∞ on M it follows that each h ∈ C (M) induces a vector field on M via the expression Xh = {·,h}, called the Hamiltonian vector field associated to the Hamiltonian function h. Hamilton’s equations ∞ z˙ = Xh(z) can be equivalently written in Poisson bracket form as f˙ = {f,h}, for any f ∈ C (M). The derivation property of the Poisson bracket implies that for any two functions f, g ∈ C∞(M), the value of the bracket {f, g}(z) at an arbitrary point z ∈ M (and therefore Xf (z) as well), depends on f only through df(z) which allows us to define a contravariant antisymmetric two–tensor B ∈ Λ2(M) by ∗ ∗ B(z)(αz, βz)= {f, g}(z), where df(z)= αz ∈ Tz M and dg(z)= βz ∈ Tz M. This tensor is called the Poisson tensor of M. The vector bundle map B♯ : T ∗M → TM naturally associated to B is defined ♯ by B(z)(αz, βz)= hαz, B (βz)i. We start by rewriting the solutions of the standard Hamilton equations in a form that we will be able to mimic in the stochastic differential equations context. All the necessary prerequisites on stochastic calculus on manifolds can be found in a short review in the appendix at the end of the paper.

Proposition 2.1 Let (M,ω) be a symplectic manifold and h ∈ C∞(M). The smooth curve γ : [0,T ] → M is an integral curve of the Hamiltonian vector ﬁeld Xh if and only if for any α ∈ Ω(M) and for any t ∈ [0,T ] t α = − dh(ω♯(α)) ◦ γ(s)ds, (2.2) Zγ|[0,t] Z0 where ω♯ : T ∗M → TM is the vector bundle isomorphism induced by ω. More generally, if M is a Poisson manifold with bracket {·, ·} then the same result holds with (2.2) replaced by

t α = − dh(B♯(α)) ◦ γ(s)ds, (2.3) Zγ|[0,t] Z0 L´azaro and Ortega: Stochastic Hamiltonian dynamical systems 4

Proof. Since in the symplectic case ω♯ = B♯, it suﬃces to prove (2.3). As (2.3) holds for any t ∈ [0,T ], we can take derivatives with respect to t on both sides and we obtain the equivalent form

hα(γ(t)), γ˙ (t)i = −hdh(γ(t)),B♯(γ(t))(α(γ(t)))i. (2.4)

Let f ∈ C∞(M) be such that df(γ(t)) = α(γ(t)). Then (2.4) can be rewritten as

hdf(γ(t)), γ˙ (t)i = −hdh(γ(t)),B♯(γ(t))(df(γ(t)))i = {f,h}(γ(t)), which is equivalent toγ ˙ (t)= Xh(γ(t)), as required. We will now introduce the stochastic Hamilton equations by mimicking in the context of Stratonovich integration the integral expressions (2.2) and (2.3). In the next definition we will use the following notation: let f : M → W be a differentiable function that takes values on the vector space W . We define the differential df : TM → W as the map given by df = p2 ◦ Tf, where Tf : TM → T W = W × W is the tangent map of f and p2 : W × W → W is the projection onto the second factor. If W = R this n i definition coincides with the usual differential. If {e1,...,en} is a basis of W and f = i=1 f ei then n i df = df ⊗ ei. i=1 P

DefinitionP 2.2 Let (M, {·, ·}) be a Poisson manifold, X : R+ × Ω → V a semimartingale that takes ∗ 1 r values on the vector space V with X0 = 0, and h : M → V a smooth function. Let {ǫ ,...,ǫ } be ∗ r i a basis of V and h = i=1 hiǫ . The Hamilton equations with stochastic component X, and Hamiltonian function h are the Stratonovich stochastic differential equation P δΓh = H(X, Γ)δX, (2.5) defined by the Stratonovich operator H(v,z) : TvV → TzM given by

r j H(v,z)(u) := hǫ ,uiXhj (z). (2.6) j=1 X ∗ ∗ ∗ ∗ The dual Stratonovich operator H (v,z) : Tz M → Tv V of H(v,z) is given by H (v,z)(αz) = ♯ −dh(z)·B (z)(αz). Hence, the results quoted in Appendix 6.4 show that for any F0 measurable random h h variable Γ0, there exists a unique semimartingale Γ such that Γ0 = Γ0 and a maximal stopping time ζh that solve (2.5), that is, for any α ∈ Ω(M),

hα, δΓhi = − hdh(B♯(α))(Γh),δXi. (2.7) Z Z h We will refer to Γ as the Hamiltonian semimartingale associated to h with initial condition Γ0.

Remark 2.3 The stochastic component X encodes the random behavior exhibited by the stochastic Hamiltonian system that we are modeling and the Hamiltonian function h speciﬁes how it embeds in its phase space. Unlike the situation encountered in the deterministic setup we allow the Hamiltonian function to be vector valued in order to accommodate higher dimensional stochastic dynamics.

Remark 2.4 The generalization of Hamilton’s equations proposed in Definition 2.2 by using a Stratonovich operator is inspired by one of the transfer principles presented in [E90] to provide stochastic versions of ordinary differential equations. This procedure can be also used to carry out a similar generalization of the equations induced by a Leibniz bracket (see [OP04]). Lázaro and Ortega: Stochastic Hamiltonian dynamical systems 5

Remark 2.5 Stratonovich versus Itôintegration: at the time of proposing the equations in Def- inition 2.2 a choice has been made, namely, we have chosen Stratonovich integration instead of Itôor other kinds of stochastic integration. The option that we took is motivated by the fact that by using Stratonovich integration, most of the geometric features underlying classical deterministic Hamiltonian mechanics are preserved in the stochastic context (see the next section). Additionally, from the mathematical point of view, this choice is the most economical one in the sense that the classical geometric ingredients of Hamiltonian mechanics plus a noise semimartingale suffice to construct the equations; had we used Itôintegration we would have had to provide a Schwartz operator (see Section 6.4) and the construction of such an object via a transfer principle like in [E90] involves the choice of a connection. The use of Itôintegration in the modeling of physical phenomena is sometimes preferred because the definition of this integral is not anticipative, that is, it does not assume any knowledge about the behavior of the system in future times. Even though we have used Stratonovich integration to write down our equations, we also share this feature because the equations in Definition 2.2 can be naturally translated to the Itôframework (see Proposition 2.8). This is a particular case of a more general fact since given any Stratonovich stochastic differential equation there always exists an equivalent Itô stochastic differential equation, in the sense that both equations have the same solutions. Note that the converse is in general not true.

2.1 Elementary properties of the stochastic Hamilton’s equations

Proposition 2.6 Let (M, {·, ·}) be a Poisson manifold, X : R+ × Ω → V a semimartingale that takes ∗ values on the vector space V with X0 =0 and h : M → V a smooth function. Let Γ0 be a F0 measurable h random variable and Γ the Hamiltonian semimartingale associated to h with initial condition Γ0. Let ζh be the corresponding maximal stopping time. Then, for any stopping time τ < ζh, the Hamiltonian semimartingale Γh satisﬁes

r τ h h h j f(Γτ ) − f(Γ0 )= {f,hj}(Γ )δX , (2.8) j=1 0 X Z j where {hj }j∈{1,...,r} and {X }j∈{1,...,r} are the components of h and X with respect to two given dual 1 r ∗ bases {e1,...,er} and {ǫ ,...,ǫ } of V and V , respectively. Expression (2.8) can be rewritten in diﬀerential notation as r h h j δf(Γ )= {f,hj}(Γ )δX . j=1 X Proof. It suﬃces to take α = df in (2.7). Indeed, by (6.5)

τ h h h hdf,δΓ i = f(Γτ ) − f(Γ0 ). Z0 At the same time

τ r τ r τ ♯ h j ♯ h h j − hdh(B (df))(Γ ),δXi = − h(dhj ⊗ ǫ (B (df)))(Γ ),δXi = h{f,hj }(Γ )ǫ ,δXi. 0 j=1 0 j=1 0 Z X Z X Z r τ h j j By the second statement in (6.5) this equals j=1 0 {f,hj}(Γ )δ hǫ ,δXi . Given that hǫ ,δXi = j j X − X0 , the equality follows. P R R R L´azaro and Ortega: Stochastic Hamiltonian dynamical systems 6

∗ ∞ Remark 2.7 Notice that if in Deﬁnition 2.2 we take V = R, h ∈ C (M), and X : R+ × Ω → R the deterministic process given by (t,ω) 7−→ t, then the stochastic Hamilton equations (2.7) reduce to

h h hα, δΓ i = hα, Xhi Γt dt. (2.9) Z Z h A straightforward application of (2.8) shows that Γt (ω) is necessarily a diﬀerentiable curve, for any ω ∈ Ω, and hence the Riemann-Stieltjes integral in the left hand side of (2.9) reduces, when evaluated at a given ω ∈ Ω, to a Riemann integral identical to the one in the left hand side of (2.3), hence proving that (2.9) reduces to the standard Hamilton equations. h h Indeed, let Γt0 (ω) ∈ M be an arbitrary point in the curve Γt (ω), let U be a coordinate patch around h 1 n 1 n h Γt0 (ω) with coordinates {x ,...,x }, and let x(t) = (x (t),...,x (t)) be the expression of Γt (ω) in these coordinates. Then by (2.8), for h ∈ R suﬃciently small, and i ∈{1,...,n},

t0+h i i i x (t0 + h) − x (t0)= {x ,h}(x(t))dt. Zt0 i Hence, by the Fundamental Theorem of Calculus, x (t) is diﬀerentiable at t0, with derivative

t0+h i 1 i i 1 i i x˙ (t0) = lim x (t0 + h) − x (t0) = lim {x ,h}(x(t))dt = {x ,h}(x(t0)), h→0 h h→0 h Zt0 ! as required.

The following proposition provides an equivalent expression of the Stochastic Hamilton equations in the Itˆoform (see Section 6.4).

Proposition 2.8 The stochastic Hamilton’s equations in Deﬁnition 2.2 admit an equivalent description using Itˆointegration by using the Schwartz operator H(v,m) : τvV → τmM naturally associated to the Hamiltonian Stratonovich operator H and that can be described as follows. Let L ∈ τvM be a second order vector and f ∈ C∞ (M) arbitrary, then

r j i j H (v,m) (L) [f]= {f,hj }(m)ǫ + {{f,hj},hi}(m)ǫ · ǫ ,L . *i,j=1 + X Moreover, expression (2.8) in the Itˆorepresentation is given by

r r τ 1 τ f Γh − f Γh = {f,h } Γh dXj + {{f,h } ,h } Γh d Xj,Xi . (2.10) τ 0 j 2 j i j=1 Z0 j,i=1 Z0 X X We will refer to H as the Hamiltonian Schwartz operator associated to h.

Proof. According to the remarks made in the Appendix 6.4, the Schwartz operator H naturally associated to H is constructed as follows. For any second order vector Lv¨ ∈ τvM associated to the acceleration of a curve v (t) in V such that v (0) = v we deﬁne H (v,m) (Lv¨) := Lm¨ (0) ∈ τmM, where m (t) is a curve in M such that m (0) = m andm ˙ (t)= H (v (t) ,m (t))v ˙ (t), for t in a neighborhood of L´azaro and Ortega: Stochastic Hamiltonian dynamical systems 7

0. Consequently, d2 d d H (v,m) (L ) [f]= f (m (t)) = hdf(m(t)), m˙ (t)i = hdf(m(t)), H(v(t),m(t))v ˙ (t)i v¨ dt2 dt dt t=0 t=0 t=0 r r d d = hǫj , v˙(t)ihdf (m(t)),X (m(t))i = hǫj , v˙(t)i{f,h }(m(t)) dt hj dt j t=0 j=1 t=0 j=1 X X r j j = hǫ , v¨(0)i{f,hj }(m)+ hǫ , v˙(0)ihd{f,hj }(m), m˙ (0)i j=1 X r r j j i = hǫ , v¨(0)i{f,hj }(m)+ hǫ , v˙(0)i hǫ , v˙(0)i{{f,hj},hi}(m) j=1 i=1 X X r j i j = {f,hj }(m)ǫ + {{f,hj},hi}(m)ǫ · ǫ ,Lv¨ . *i,j=1 + X ∗ In order to establish (2.10) we need to calculate H (v,m) (d2f(m)) for a second order form d2f(m) ∈ ∗ ∞ ∗ τmM at m ∈ M, f ∈ C (M). Since H (v,m) (d2f(m)) is fully characterized by its action on elements of the form Lv¨ ∈ τvV for some curve v (t) in V such that v (0) = v, we have ∗ hH (v,m) (d2f(m)),Lv¨i = hd2f(m), H (v,m) (Lv¨)i = H (v,m) (Lv¨) [f] r j i j = {f,hj}(m)ǫ + {{f,hj},hi}(m)ǫ · ǫ ,Lv¨ . *i,j=1 + X ∗ r j i j Consequently, H (v,m) (d2f(m)) = i,j=1{f,hj }(m)ǫ + {{f,hj},hi}(m)ǫ · ǫ . Hence, if Γ is the Hamiltonian semimartingale associated to h with initial condition Γ , τ < ζh is h P 0 any stopping time, and f ∈ C∞(M), we have by (6.5) and (6.6) τ τ h h h ∗ h f Γτ − f Γ0 = d2f, dΓ = H X, Γ (d2f), dX Z0 Z0 r τ r τ h j h i j = {f,hj } Γ ǫ , dX + {{f,hj} ,hi} Γ ǫ · ǫ , dX j=1 0 j,i=1 0 X Z X Z r r τ 1 τ = {f,h } Γh dXj + {{f,h } ,h } Γh d Xi,Xj . j 2 j i j=1 Z0 j,i=1 Z0 X X Proposition 2.9 (Preservation of the symplectic leaves by Hamiltonian semimartingales) In the setup of Deﬁnition 2.2, let L be a symplectic leaf of (M,ω) and Γh a Hamiltonian semimartingale with initial condition Γ0(ω) = Z0, where Z0 is a random variable such that Z0(ω) ∈ L for all ω ∈ Ω. h h Then, for any stopping time τ < ζ we have that Γτ ∈ L. Proof. Expression (2.6) shows that for any z ∈ L, the Stratonovich operator H(v,z) takes values in the characteristic distribution associated to the Poisson structure (M, {·, ·}), that is, in the tangent space T L of L. Consequently, H induces another Stratonovich operator HL(v,z) : TvV → TzL, v ∈ V , z ∈ L, obtained from H by restriction of its range. It is clear that if i : L ֒→ M is the inclusion then

∗ ∗ ∗ HL(v,z) ◦ Tz i = H (v,z). (2.11) h Let ΓL be the semimartingale in L that is a solution of the Stratonovich stochastic diﬀerential equation h h δΓL = HL(X, ΓL)δX (2.12) L´azaro and Ortega: Stochastic Hamiltonian dynamical systems 8

h with initial condition Γ0. We now show that Γ := i ◦ ΓL is a solution of δΓ= H(X, Γ)δX.

The uniqueness of the solution of a stochastic diﬀerential equation will guarantee in that situation that Γh necessarily coincides with Γ, hence proving the statement. Indeed, for any α ∈ Ω(M),

h ∗ h hα, δΓi = hα, δ(i ◦ ΓL)i = hT i · α, δΓLi. Z Z Z h ∗ Since ΓL satisﬁes (2.12) and T i · α ∈ Ω(L), by (2.11) this equals

∗ h ∗ ∗ h ∗ hHL(X, ΓL)(T i · α),δXi = hH (X,i ◦ ΓL)(α),δXi = hH (X, Γ)(α),δXi, Z Z Z that is, δΓ= H(X, Γ)δX, as required.

Proposition 2.10 (The stochastic Hamilton equations in Darboux-Weinstein coordinates) Let (M, {·, ·}) be a Poisson manifold and Γh be a solution of the Hamilton equations (2.5) with initial condition x0 ∈ M. There exists an open neighborhood U of x0 in M and a stopping time τU such h that Γt (ω) ∈ U, for any ω ∈ Ω and any t ≤ τU (ω). Moreover, U admits local Darboux coordinates 1 n (q ,...,q ,p1,...,pn,z1,...,zl) in which (2.8) takes the form

r τ i h i h ∂hj j q (Γτ ) − q (Γ0 ) = δX , ∂pi j=1 0 X Z r τ ∂h p (Γh) − p (Γh) = − j δXj, i τ i 0 ∂qi j=1 0 X Z r τ h h j zi(Γτ ) − zi(Γ0 ) = {zi,hj }T δX , j=1 0 X Z where {·, ·}T is the transverse Poisson structure of (M, {·, ·}) at x0.

Proof. Let U be an open neighborhood of x0 in M for which Darboux coordinates can be chosen. h c Define τU = inft≥0{Γt ∈ U } (τU is the exit time of U). It is a standard fact in the theory of stochastic processes that τU is a stopping time. The proposition follows by writing (2.8) for the Darboux-Weinstein 1 n coordinate functions (q ,...,q ,p1,...,pn,z1,...,zl). Let ζ : M × Ω → [0, ∞] be the map such that, for any z ∈ M, ζ (z) is the maximal stopping time associated to the solution of the stochastic Hamilton equations (2.5) with initial condition Γ0 = z a.s.. Let F be the flow of (2.5), that is, for any z ∈ M, F (z):[0, ζ (z)] → M is the solution semimartingale of (2.5) with initial condition z. The map z ∈ M 7−→ Ft(z,ω) ∈ M is a local diffeomorphism of M, for each t ≥ 0 and almost all ω ∈ Ω in which this map is defined (see [IW89]). In the following result, we show that, in the symplectic context, Hamiltonian flows preserve the symplectic form and hence the associated volume form θ = ω ∧ ...n ∧ ω. This has already been shown for Hamiltonian diffusions (see Example 3.1) by Bismut [B81].

Theorem 2.11 (Stochastic Liouville’s Theorem) Let (M,ω) be a symplectic manifold, X : R+ × Ω → V ∗ a semimartingale, and h : M → V ∗ a Hamiltonian function. Let F be the associated Hamilto- nian ﬂow. Then, for any z ∈ M and any (t, η) ∈ [0, ζ (z)],

∗ Ft (z, η) ω = ω. L´azaro and Ortega: Stochastic Hamiltonian dynamical systems 9

Proof. By [K81, Theorem 3.3] (see also [W80]), given an arbitrary form α ∈ Ωk (M) and z ∈ M, the process F (z)∗ α satisﬁes the following stochastic diﬀerential equation:

r F (z)∗ α = α (z)+ F (z)∗ £ α δXj. Xhj j=1 X Z In particular, if α = ω then £ ω = 0 for any j ∈{1, ..., r}, and hence the result follows. Xhj

2.2 Conserved quantities and stability Conservation laws in Hamiltonian mechanics are extremely important since they make easier the integration of the systems that have them and, in some instances, provide qualitative information about the dynamics. A particular case of this is their use in concluding the nonlinear stability of certain equilibrium solutions using Dirichlet type criteria that we will generalize to the stochastic setup using the following deﬁnitions.

Definition 2.12 A function f ∈ C∞ (M) is said to be a strongly (respectively, weakly) conserved quantity of the stochastic Hamiltonian system associated to h : M → V ∗ if for any solution Γh of the h h h h stochastic Hamilton equations (2.5) we have that f Γ = f Γ0 (respectively, E[f Γτ ]= E[f Γ0 ], for any stopping time τ). Notice that strongly conserved quantities are obviously weakly conserved and that the two definitions coincide for deterministic systems with the standard definition of conserved quantity. The following result provides in the stochastic setup an analogue of the classical characterization of the conserved quantities in terms of Poisson involution properties.

Proposition 2.13 Let (M, {·, ·}) be a Poisson manifold, X : R+ × Ω → V a semimartingale that takes ∗ ∞ values on the vector space V such that X0 =0, and h : M → V and f ∈ C (M) two smooth functions. If {f,hj } = 0 for every component hj of h then f is a strongly conserved quantity of the stochastic Hamilton equations (2.5). r j i j Conversely, suppose that the semimartingale X = j=1 X ǫj is such that X ,X =0 if i 6= j. If f is a strongly conserved quantity then {f,h } =0, for any j ∈{1, ..., r} such that Xj ,Xj is an strictly j P increasing process at 0. The last condition means that there exists A ∈F and δ > 0 with P (A) > 0 such j j j j that for any t<δ and ω ∈ A we have [X ,X ]t(ω) > [X ,X ]0(ω), for all j ∈{1, ..., r}.

h h Proof. Let Γ be the Hamiltonian semimartingale associated to h with initial condition Γ0 . As we saw in (2.10),

r r 1 f Γh = f Γh + {f,h } Γh dXj + {{f,h } ,h } Γh d Xi,Xj . (2.13) 0 j 2 j i j=1 Z j,i=1 Z X X If {f,hj } = 0 for every component hj of h then all the integrals in the previous expression vanish and h h therefore f Γ = f Γ0 which implies that f is a strongly conserved quantity of the Hamiltonian stochastic equations associated to h. Conversely, suppose now that f is a strongly conserved quantity. This implies that for any initial h h condition Γ0 , the semimartingale f Γ is actually time independent and hence of finite variation. Equivalently, the (unique) decomposition of f Γh into two processes, one of finite variation plus a local martingale, only has the first term. In order to isolate the local martingale term of f Γh recall first that the quadratic variations Xi,Xj have finite variation and that the integral with respect to a finite variation process has finite variation (see [LeG97, Proposition 4.3]). Consequently, the last Lázaro and Ortega: Stochastic Hamiltonian dynamical systems 10 summand in (2.13) has finite variation. As to the second summand, let M j and Aj , j =1,...,r, local martingales and finite variation processes, respectively, such that Xj = Aj + M j . Then,

h j h j h j {f,hj} Γ dX = {f,hj } Γ dM + {f,hj} Γ dA . Z Z Z h j h j Given that for each j, {f,hj } Γ dA is a ﬁnite variation process and {f,hj} Γ dM is a local martingale (see [P90, Theorem 29, page 128]) we conclude that Z := r {f,h } Γh dM j is the R jR=1 j local martingale term of f Γh and hence equal to zero. P R 2 We notice now that any continuous local martingale Z : R+ ×Ω → R is also a local L (Ω)-martingale. n τ n 2 Indeed, consider the sequence of stopping times τ = {inf t ≥ 0 | |Zt| = n}, n ∈ N. Then E Z t ≤ 2 2 τ n 2 τ n 2 h τ n τ ni E n = n , for all t ∈ R+. Hence, Z ∈ L (Ω) for any n. In addition, E Z t = E Z ,Z t (see [P90, Corollary 3, page 73]). On the other hand by Proposition 5.5, h i τ n r r τ n h j h j Z = {f,hj} Γ dM = 1 n {f,hj } Γ dM .   [0,τ ] j=1 Z j=1 Z X X   Thus, by [P90, Theorem 29, page 75] and the hypothesis Xi,Xj = 0 if i 6= j,

r 2 τ n τ n τ n h j h i E Z = E Z ,Z = E 1[0,τ n] {f,hj} Γ dM , 1[0,τ n] {f,hi} Γ dM t t j,i=1 t hh i i X Z Z r h j i = E 1[0,τ n] ({f,hj }{f,hi}) Γ d M ,M j,i=1 t X Z r h j i = E 1[0,τ n] ({f,hj }{f,hi}) Γ d X ,X j,i=1 t X Z r 2 h j j = E 1[0,τ n] {f,hj } Γ d X ,X . j=1 Z t X j j 2 h j j Since X ,X is an increasing process of ﬁnite variation then 1[0,τ n] {f,hj} Γ d X ,X is a Riemann-Stieltjes integral and hence for any ω ∈ Ω R

2 h j j 2 h j j 1[0,τ n] {f,hj } Γ d X ,X (ω)= 1[0,τ n(ω)] {f,hj} Γ (ω) d X ,X (ω) . Z Z j j As X ,X (ω) is an increasing function of t ∈ R+, then for any j ∈{1,...,r} 2 h j j E 1[0,τ n] {f,hj } Γ d X ,X ≥ 0. (2.14) Z τ n 2 Additionally, since E Z t = 0, we necessarily have that the inequality in (2.14) is actually an equality. Hence, h i t 2 h j j 1[0,τ n] {f,hj } Γ d X ,X =0. (2.15) Z0 Suppose now that [Xj ,Xj] is strictly increasing at 0 for a particular j. Hence, there exists A ∈ j j j j F with P (A) > 0, and δ > 0 such that [X ,X ]t (ω) > [X ,X ]0 (ω) for any t < δ. Take now a L´azaro and Ortega: Stochastic Hamiltonian dynamical systems 11

ﬁxed ω ∈ A. Since τ n → ∞ a.s., we can take n large enough to ensure that τ n (ω) > t, where j j t ∈ [0,δ). Thus, we may suppose that 1[0,τ n] (t,ω)=1. As X ,X (ω) is an strictly increasing process t at zero {f,h }2 Γh (ω) d Xj,Xj (ω) > 0 unless {f,h }2 Γh (ω) = 0 in a neighborhood [0, δ ) of 0 j j ω 0 containedR in [0,δ). In principle δω > 0 might depend on ω ∈ A, so the values of t ∈ [0,δ) for which 2 h e {f,hj } Γt (ω) = 0 for any ω ∈ A are those verifying 0 ≤ t ≤ infω∈A δω. In any case (2.15) allows us 2 h e h to conclude that {f,hj} Γ0 (ω) = 0 for any ω ∈ A. Finally, consider any Γ solution to the Stochastic h Hamilton equations with constant initial condition Γ0 = m ∈ M an arbitrarye point. Then, for any ω ∈ A, 2 h 2 0= {f,hj } Γ0 (ω) = {f,hj } (m) .

Since m ∈ M is arbitrary we can conclude that {f,hj} = 0. We now use the conserved quantities of a system in order to formulate suﬃcient Dirichlet type stability criteria. Even though the statements that follow are enounced for processes that are not necessarily Hamiltonian, it is for these systems that the criteria are potentially most useful. We start by spelling out the kind of nonlinear stability that we are after.

Deﬁnition 2.14 Let M be a manifold and let

δΓ= e(X, Γ)δX (2.16) be a Stratonovich stochastic diﬀerential equation whose solutions Γ : R × Ω → M take values on M. s,x s,x Given x ∈ M and s ∈ R, denote by Γ the unique solution of (2.16) such that Γs (ω) = x, for all ω ∈ Ω. Suppose that the point z0 ∈ M is an equilibrium of (2.16), that is, the constant process Γt(ω) := z0, for all t ∈ R and ω ∈ Ω, is a solution of (2.16). Then we say that the equilibrium z0 is

(i) Almost surely (Lyapunov) stable when for any open neighborhood U of z0 there exists another 0,z neighborhood V ⊂ U of z0 such that for any z ∈ V we have Γ ⊂ U, a.s. (ii) Stable in probability. For any s ≥ 0 and ǫ> 0

s,x lim P sup d (Γt ,z0) >ǫ =0, x→z 0 t>s where d : M × M → R is any distance function that generates the manifold topology of M.

Theorem 2.15 (Stochastic Dirichlet’s Criterion) Suppose that we are in the setup of the previ- ∞ ous deﬁnition and assume that there exists a function f ∈ C (M) such that df(z0)=0 and that the 2 quadratic form d f(z0) is (positive or negative) deﬁnite. If f is a strongly (respectively, weakly) conserved quantity for the solutions of (2.16) then the equilibrium z0 is almost surely stable (respectively, stable in probability).

Proof. Since the stability of the equilibrium z0 is a local statement, we can work in a chart of M around z0 with coordinates (x1,...,xn) in which z0 is modeled by the origin. Moreover, using the Morse lemma and the hypotheses on the function f, and assuming without loss of generality that f(z0) = 0, 2 2 we choose the coordinates (x1,...,xn) so that f(x1,...,xn)= x1 + · · · + xn. Hence, in the definition of stability in probability, we can use the distance function d(x, z0)= f(x). Suppose now that f is a strongly conserved quantity and let U be an open neighborhood of z0. Let r > 0 be such that V := f −1([0, r]) ⊂ U. Let z ∈ V with f(z) = r′. As f is a strongly conserved quantity f(Γ0,z)= r′ ≤ r and hence Γ0,z ⊂ U, as required. In order to study the case in which f is a weakly conserved quantity, let ǫ > 0 and let Uǫ be the s,x ball of radius ǫ around z0. Then, for any x ∈ Uǫ and s ∈ R+, let τUǫ be the first exit time of Γ with Lázaro and Ortega: Stochastic Hamiltonian dynamical systems 12

s,x respect to Uǫ. Notice ﬁrst that if ω ∈ Ω belongs to the set {ω ∈ Ω | sup0≤s ǫ} = {ω ∈ s,x 2 s,x τUǫ Ω | sup0≤sǫ }, then τUǫ (ω) ≤ t and hence the stopped process (Γ ) satisﬁes that

f ((Γs,x)τUǫ (ω)) = f Γs,x (ω) = ǫ2, t τUǫ (ω) for those values of ω. This ensures that

2 s,x τU ǫ 1 s,x ≤ f ((Γ ) ǫ ) . {ω∈Ω|sup0≤sǫ} t Taking expectations in both sides of this inequality we obtain

E[f ((Γs,x)τUǫ )] P sup d (Γs,x,z ) >ǫ ≤ t . t 0 ǫ2 0≤sǫ ≤ . (2.17) t 0 ǫ2 0≤s

Taking the limit x → z0 in this expression and recalling that f(z0) = 0, the result follows. A careful inspection of the proof that we just carried out reveals that in order for (2.17) to hold, it would suﬃce to have E[f (Γτ )] ≤ E[f (Γ0)], for any stopping time τ and any solution Γ, instead of the equality guaranteed by the weak conservation condition. This motivates the next deﬁnition.

Deﬁnition 2.16 Suppose that we are in the setup of Deﬁnition 2.14. Let U be an open neighborhood of the equilibrium z0 and let V : U → R be a continuous function. We say that V is a Lyapunov function for the equilibrium z0 if V (z0)=0, V (z) > 0 for any z ∈ U \{z0}, and

E[V (Γτ )] ≤ E[V (Γ0)], (2.18) for any stopping time τ and any solution Γ of (2.16).

This definition generalizes to the stochastic context the standard notion of Lyapunov function that one encounters in dynamical systems theory. If (2.16) is the stochastic differential equation associated to an Itôdiffusion and the Lyapunov function is twice differentiable, the inequality (2.18) can be ensured by requiring that A[V ](z) ≤ 0, for any z ∈ U \{z0}, where A is the infinitesimal generator of the diffusion, and by using Dynkin’s formula.

Theorem 2.17 (Stochastic Lyapunov’s Theorem) Let z0 ∈ M be an equilibrium solution of the stochastic diﬀerential equation (2.16) and let V : U → R be a continuous Lyapunov function for z0. Then z0 is stable in probability.

Proof. Let Uǫ be the ball of radius ǫ around z0 and let Vǫ := infx∈U\Uǫ V (x). Using the same notation as in the previous theorem we denote, for any x ∈ Uǫ and s ∈ R+, τUǫ as the first exit time of s,x Γ with respect to Uǫ. Using the same approach as above we notice that if ω ∈ Ω belongs to the set s,x s,x τUǫ {ω ∈ Ω | sup0≤s ǫ}, then τUǫ (ω) ≤ t and hence the stopped process (Γ ) satisfies that s,x τUǫ s,x V ((Γ ) (ω)) = V Γ (ω) ≥ Vǫ, t τUǫ (ω) Lázaro and Ortega: Stochastic Hamiltonian dynamical systems 13

s,x for those values of ω, since Γ (ω) belongs to the boundary of Uǫ. This ensures that τUǫ (ω)

s,x τU V 1 s,x ≤ V ((Γ ) ǫ ) . ǫ {ω∈Ω|sup0≤sǫ} t Taking expectations in both sides of this inequality we obtain E[V ((Γs,x)τUǫ )] P sup d (Γs,x,z ) >ǫ ≤ t . t 0 V 0≤s

s,x s,x τUǫ E V Γ s,x E[V ((Γ ) )] τUǫ ∧t E[V (Γ )] V (x) t = ≤ s = . Vǫ h Vǫ i Vǫ Vǫ We can therefore conclude that V (x) P sup d (Γs,x,z ) >ǫ ≤ . t 0 V 0≤s

Taking the limit x → z0 in this expression and recalling that V (z0) = 0, the result follows. Remark 2.18 This theorem has been proved by Gihman [G66] and Hasminskii [Ha80] for Itˆodiﬀusions.

3 Examples 3.1 Stochastic perturbation of a Hamiltonian mechanical system and Bis- mut’s Hamiltonian diﬀusions

∞ Let (M, {·, ·}) be a Poisson manifold and hj ∈ C (M), j = 0, ..., r, smooth functions. Let h : M −→ r+1 R be the Hamiltonian function m 7−→ (h0 (m) ,...,hr (m)), and consider the semimartingale X : r+1 1 r j R+ × Ω → R given by (t,ω) 7−→ t,Bt (ω) ,...,Bt (ω) , where B , j = 1, ..., r, are r-independent Brownian motions. L´evy’s characterization of Brownian motion shows (see for instance [P90, Theorem j i ji 40, page 87]) that B ,B t = tδ . In this setup, the equation (2.8) reads τ r τ h h h h j f Γτ − f Γ0 = {f,h0} Γ dt + {f,hj } Γ δB (3.1) Z0 j=1 Z0 X for any f ∈ C∞ (M). According to (2.10), the equivalent Itˆoversion of this equation is

τ r τ τ h h h h j h f Γτ − f Γ0 = {f,h0} Γ dt + {f,hj } Γ dB + {{f,hj} ,hj } Γ dt. Z0 j=1 Z0 Z0 X Equation (3.1) may be interpreted as a stochastic perturbation of the classical Hamilton equations associated to h0, that is, d(f ◦ γ) (t)= {f,h } (γ (t)) . dt 0 by the r Brownian motions Bj . These equations have been studied by Bismut in [B81] in the particular case in which the Poisson manifold (M, {·, ·}) is just the symplectic Euclidean space R2n with the canonical symplectic form. He refers to these particular processes as Hamiltonian diffusions. If we apply Proposition 2.13 to the stochastic Hamiltonian system (3.1), we obtain a generalization to Poisson manifolds of a result originally formulated by Bismut (see [B81, Théorèmes 4.1 and 4.2, page 231]) for Hamiltonian diffusions. See also [M99]. Lázaro and Ortega: Stochastic Hamiltonian dynamical systems 14

Proposition 3.1 Consider the stochastic Hamiltonian system introduced in (3.1). Then f ∈ C∞ (M) is a conserved quantity if and only if

{f,h0} = {f,h1} = . . . = {f,hr} =0. (3.2) Proof. If (3.2) holds then f is clearly a conserved quantity by Proposition 2.13. Conversely, notice that as [Bi,Bj ]= tδij , i, j ∈{1,...,r}, and X0(t,ω)= t is a finite variation process then [Xi,Xj ]=0 for any i, j ∈ {0, 1,...,r} such that i 6= j. Consequently, by Proposition 2.13, if f is a conserved quantity then {f,h1} = . . . = {f,hr} =0. (3.3) Moreover, (3.1) reduces to τ h {f,h0} Γ dt =0, Z0 h h for any Hamiltonian semimartingale Γ and any stopping time τ ≤ ζ . Suppose that {f,h0} (m0) > 0 for some m0 ∈ M. By continuity there exists a compact neighborhood U of m0 such that {f,h0}|U > 0. h h Take Γ the Hamiltonian semimartingale with initial condition Γ0 = m0, and let ξ be the first exit time of U for Γh. Then, defining τ := ξ ∧ ζ, τ τ h {f,h0} Γ dt ≥ min {{f,h0} (m) | m ∈ U} dt > 0, Z0 Z0 which contradicts (3.3). Therefore, {f,h0} = 0 also, as required. Remark 3.2 Notice that, unlike what happens for standard deterministic Hamiltonian systems, the energy h0 of a Hamiltonian diffusion does not need to be conserved if the other components of the Hamiltonian are not involution with h0. This is a general fact about stochastic Hamiltonian systems that makes them useful in the modeling of dissipative phenomena. We see more of this in the next example.

3.2 Integrable stochastic Hamiltonian dynamical systems.

∗ Let (M,ω)bea2n-dimensional manifold, X : R+ × Ω → V a semimartingale, and h : M → V such r i 1 r ∗ that h = i=1 hiǫ , with ǫ , ..., ǫ a basis of V . Let H be the associated Stratonovich operator in (2.6). P ∞ Suppose that there exists a family of functions {fr+1, ..., fn} ⊂ C (M) such that the n-functions ∞ {f1 := h1, ..., fr := hr,fr+1, ..., fn} ⊂ C (M) are in Poisson involution, that is, {fi,fj} = 0, for any i, j ∈{1, ..., n}. Moreover, assume that F := (f1, ..., fn) satisfies the hypotheses of the Liouville-Arnold Theorem [Ar89]: F has compact and connected fibers and its components are independent. In this setup, we will say that the stochastic Hamiltonian dynamical system associated to H is integrable. As it was already the case for standard (Liouville-Arnold) integrable systems, there is a symplec- n n n i tomorphism that takes (M,ω) to T × R , i=1 dθ ∧ dIi and for which F ≡ F (I1, ..., In). In particular, in the action-angle coordinates I , ..., I ,θ1, ..., θn , h ≡ h (I , ..., I ) with j ∈ {1, ..., r}. In 1 P n j j 1 n other words, the components of the Hamiltonian function depend only on the actions I := (I , ..., In). 1 Therefore, for any random variable Γ0 and any i ∈{1, ..., n} r j Ii (Γ) − Ii (Γ0)= {Ii,hj (I)} (Γ) δX = 0 (3.4a) j=1 X Z r r i i i j ∂hj j θ (Γ) − θ (Γ0)= θ ,hj (I) (Γ) δX = (Γ) δX . (3.4b) ∂Ii j=1 Z j=1 Z X X Lázaro and Ortega: Stochastic Hamiltonian dynamical systems 15

Consequently, the tori determined by fixing I = constant are left invariant by the stochastic flow associated to (3.4). In particular, as the paths of the solutions are contained in compact sets, the stochastic flow is defined for any time and the flow is complete. Moreover, the restriction of this stochastic differential equation to the torus given by say, I0, yields the solution

r i i j θ (Γ) − θ (Γ0)= ωj (I0) X , (3.5) j=1 X where ω (I ) := ∂hj (I ) and where we have assumed that X = 0. Expression (3.5) clearly resembles j 0 ∂Ii 0 0 the integration that can be carried out for deterministic integrable systems. Additionally, the Haar measure dθ1 ∧ ... ∧ dθn on each invariant torus is left invariant by the stochastic flow (see Theorem 2.11 and [L06]). Therefore, if we can ensure that there exists a unique invariant measure µ (for instance, if (3.5) defines a non-degenerate diffusion on the torus Tn, the invariant measure is unique up to a multiplicative constant by the compactness of Tn (see [IW89, Proposition 4.5])) then µ coincides necessarily with the Haar measure.

3.3 The Langevin equation and viscous damping Hamiltonian stochastic differential equations can be used to model dissipation phenomena. The simplest example in this context is the damping force experienced by a particle in motion in a viscous fluid. This dissipative phenomenon is usually modeled using a force in Newton’s second law that depends linearly on the velocity of the particle (see for instance [LL76, §25]). The standard microscopic description of this motion is carried out using the Langevin stochastic differential equation (also called the Orstein- Uhlenbeck equation) that says that the velocityq ˙(t) of the particle with mass m is a stochastic process that solves the stochastic differential equation

m dq˙(t)= −λq˙(t)dt + bdBt, (3.6) where λ> 0 is the damping coefficient, b is a constant, and Bt is a Brownian motion. A common physical interpretation for this equation (see [CH06]) is that the Brownian motion models random instantaneous bursts of momentum that are added to the particle by collision with lighter particles, while the mean effect of the collisions is the slowing down of the particle. This fact is mathematically described by saying that the expected value qe := E[q] of the process q determined by (3.6) satisfies the ordinary differential equationq ë = −λq˙e. Even though this description is accurate it is not fully satisfactory given that it does not provide any information about the mechanism that links the presence of the Brownian perturbation to the emergence of damping in the equation. In order for the physical explanation to be complete, a relation between the coefficients b and λ should be provided in such a way that the damping vanishes when the Brownian collisions disappear, that is, λ = 0 when b = 0. We now show that the motion of a particle of mass m in one dimension subjected to viscous damping with coefficient λ and to a harmonic potential with Hooke constant k is a Hamiltonian stochastic differential equation. More explicitly, we will give a stochastic Hamiltonian system such that the expected value qe of its solution semimartingales satisfies the ordinary differential equation of the damped harmonic oscillator, that is, m që(t)= −λq˙e(t) − kqe(t). This description provides a mathematical mechanism by which the stochastic perturbations in the system generate an average damping. 2 Consider R with its canonical symplectic form and let X : R+ × Ω → R be the real semimartingale 2 given by Xt (ω) = (t + νBt (ω)) with ν ∈ R and Bt a Brownian motion. Let now h : R → R be the Lázaro and Ortega: Stochastic Hamiltonian dynamical systems 16

1 2 1 2 energy of a harmonic oscillator, that is, h (q,p) := 2m p + 2 ρq . By (2.10), the solution semimartingales Γh of the Hamiltonian stochastic equations associated to h and X satisfy 1 ν q Γh − q Γh = 2p Γh − ν2ρq Γh dt + p Γh dB , (3.7) 0 2m t t m t t Z Z ρ p Γh − p Γh = − ν2p Γh +2mq Γh dt − νρ q Γh dB . (3.8) 0 2m t t t t Z Z h h Given that E p Γt dBt = E q Γt dBt = 0, if we denote

R R h h qe (t) := E q Γt , pe (t) := E p Γt ,

Fubini’s Theorem guarantees that

1 ν2ρ ν2ρ q˙ (t)= p (t) − q (t) andp ˙ (t)= − p (t) − ρq (t) . (3.9) e m e 2m e e 2m e e From the ﬁrst of these equations we obtain that

ν2ρ p (t)= mq˙ + q e e 2 e whose time derivative is ν2ρ p˙ (t)= mq¨ + q˙ . e e 2 e These two equations substituted in the second equation of (4.8) yield

ν4ρ mq¨ (t)= −ν2ρq˙ (t) − ρ +1 q (t) , (3.10) e e 4m e that is, the expected value of the position of the Hamiltonian semimartingale Γh associated to h and X satisﬁes the diﬀerential equation of a damped harmonic oscillator (5.8) with constants

ν4ρ λ = ν2ρ and k = ρ +1 . 4m Notice that the dependence of the damping and elastic constants on the coeﬃcients of the system is physically reasonable. For instance, we see that the more intense the stochastic perturbation is, that is, the higher ν is, the stronger the damping becomes (λ = ν2ρ increases). In particular, if there is no stochastic perturbation, that is, if ν = 0, then the damping vanishes, k = ρ and (4.10) becomes the diﬀerential equation of a free harmonic oscillator of mass m and elastic constant ρ.

The stability of the resting solution. It is easy to see that the constant process Γt(ω)=(0, 0), for all t ∈ R and ω ∈ Ω is an equilibrium solution of (3.7) and (3.8). One can show using the stochastic Dirichlet’s criterion (Theorem 2.15) that this equilibrium is almost surely Lyapunov stable since the Hamiltonian function h is a strongly conserved quantity (by (2.8)) that exhibits a critical point at the origin with definite Hessian. The Langevin equation. In the previous paragraphs we succeeded in providing a microscopic Hamil- tonian description of the harmonic oscillator subjected to Brownian perturbations whose macroscopic counterpart via expectations yields the equations of the damped harmonic oscillator. In view of this, is such a stochastic Hamiltonian description available for the pure Langevin equation (3.6)? The an- swer is no. More specifically, it can be easily shown (proceed by contradiction) that (3.6) cannot Lázaro and Ortega: Stochastic Hamiltonian dynamical systems 17 be written as a stochastic Hamiltonian differential equation on R2 with its canonical symplectic form with a noise semimartingale of the form Xt(ω) = (f0(t,Bt),f1(t,Bt)) and a Hamiltonian function ∞ h(q,p) = (h0(q,p),h1(q,p)), f0,f1,h0,h1 ∈ C (R). Nevertheless, if we put aside for a moment the stochastic Hamiltonian category and we use Itôintegration, the Langevin equation can still be written in phase space, that is, dqt = vtdt, dvt = −λvtdt + bdBt, (3.11) as a stochastic perturbation of a deterministic system, namely, a free particle whose evolution is given by the differential equations dqt = vtdt and dvt =0. (3.12) 1 2 2 Let u ,u be global coordinates on R associated to the canonical basis {e1,e2} and consider the global basis d ui, d ui · d uj of τ ∗R2. Define a dual Schwartz operator S∗ (x, (q, v)) : τ ∗ R2 −→ τ ∗R2 2 2 2 i,j=1,2 (q,v) x characterized by the relations 1 2 2 2 d2q 7−→ vd2u , d2v 7−→ bd2u − λv d2u · d2u ,

2 2 2 where (q, v) ∈ R is an arbitrary point in phase space and x ∈ R . If X : R+ × Ω → R is such that X (t,ω) = (t,bBt (ω)), for any (t,ω) ∈ R+ × Ω, it is immediate to see that the Itˆoequations associated to S∗ and X are (3.11). Moreover, if we set b = 0, that is, we switch oﬀ the Brownian perturbation then we recover (3.12), as required.

3.4 Brownian motions on manifolds The mathematical formulation of Brownian motions (or Wiener processes) on manifolds has been the subject of much research and it is a central topic in the study of stochastic processes on manifolds (see [IW89, Chapter 5], [E89, Chapter V], and references therein for a good general review of this subject). In the following paragraphs we show that Brownian motions can be defined in a particularly simple way using the stochastic Hamilton equations introduced in Definition 2.2. More specifically we will show that Brownian motions on manifolds can be obtained as the projections onto the base space of very simple Hamiltonian stochastic semimartingales defined on the cotangent bundle of the manifold or of its orthonormal frame bundle, depending on the availability or not of a parallelization for the manifold in question. We will first present the case in which the manifold in question is parallelizable or, equivalently, when the coframe bundle on the manifold admits a global section, for the construction is particularly simple in this situation. The parallelizability hypothesis is verified by many important examples. For instance, any Lie group is parallelizable; the spheres S1, S3, and S7 are parallelizable too. At the end of the section we describe the general case. The notion of manifold valued Brownian motion that we will use is the following. A M-valued process Γ is called a Brownian motion on (M,g), with g a Riemannian metric on M, whenever Γ is continuous and adapted and for every f ∈ C∞(M)

1 f(Γ) − f(Γ ) − ∆ f(Γ)dt 0 2 M Z is a local martingale. We recall that the Laplacian ∆M (f) is deﬁned as ∆M (f) = Tr (Hess f), for any f ∈ C∞ (M), where Hess f := ∇(∇f), with ∇ : X(M) × X(M) → X(M), the Levi-Civita connection of g. Hess f is a symmetric (0, 2)-tensor such that for any X, Y ∈ X(M),

Hess f(X, Y )= X [g(grad f, Y )] − g(grad f, ∇X Y ). (3.13) L´azaro and Ortega: Stochastic Hamiltonian dynamical systems 18

Brownian motions on parallelizable manifolds. Suppose that the n-dimensional manifold (M,g) is parallelizable and let {Y1, ..., Yn} be a family of vector ﬁelds such that for each m ∈ M, {Y1(m), ..., Yn(m)} forms a basis of TmM (a parallelization). Applying the Gram-Schmidt orthonormalization procedure if necessary, we may suppose that this parallelization is orthonormal, that is, g (Yi, Yj ) = δij , for any i, j =1, ..., n. Using this structure we are going to construct a stochastic Hamiltonian system on the cotangent bundle T ∗M of M, endowed with its canonical symplectic structure, and we will show that the projection of the solution semimartingales of this system onto M are M-valued Brownian motions in the sense spec- n+1 1 n iﬁed above. Let X : R+ × Ω → R be the semimartingale given by X(t,ω) := (t,Bt (ω),...,Bt (ω)), j ∗ n+1 where B , j =1, ..., n, are n-independent Brownian motions and let h = (h0,h1,...,hn) : T M → R be the function whose components are given by

∗ ∗ h0 : T M −→ R hj : T M −→ R 1 n and (3.14) αm 7−→ − 2 j=1hαm, ∇Yj Yj (m)i αm 7−→ hαm, Yj (m)i. We will now study the projectionP onto M of Hamiltonian semimartingales Γh that have X as stochastic component and h as Hamiltonian function and will prove that they are M-valued Brownian motions. In order to do so we will be particularly interested in the projectable functions f of T ∗M, that is, the functions f ∈ C∞(T ∗M) that can be written as f = f ◦ π with f ∈ C∞(M) and π : T ∗M → M the canonical projection. We start by proving that for any projectable function f = f ◦ π ∈ C∞(T ∗M)

n 1 {f,h } = g grad f, − ∇Y Yj and {f,hj} = g grad f, Yj , (3.15) 0  2 j  j=1 X   and where {·, ·} is the Poisson bracket associated to the canonical symplectic form on T ∗M. Indeed, let ∗ 1 n i i U a Darboux patch for T M with associated coordinates q ,...,q ,p1,...,pn such that {q ,pj } = δj . There exists functions f k ∈ C∞(π(U)), with k, j ∈{1,...,n} such that the vector ﬁelds may be locally j n k ∂ n k written as Yj = k=1 fj ∂qk . Moreover, hj (q,p)= k=1 fj (q) pk and

n n n n P P ∂(f ◦ π) ∂f {f,h } = f ◦ π, f kp = f k f ◦ π,p = f k qi,p = f kδi j j k j k j ∂qi k j k ∂qi ( k=1 ) k=1 k,i=1 k,i=1 X X X X = Yj [f] ◦ π = g grad f, Yj ◦ π, as required. The ﬁrst equality in (3.15) is proved analogously. Notice that the formula that we just proved shows that if f is projectable then so is {f,hj}, with j ∈{1,...,n}. Hence, using (3.15) again and (3.13) we obtain that

{{f,hj } ,hj } = Yj g grad f, Yj ◦ π = Hess f (Yj , Yj ) ◦ π + g grad f, ∇Yj Yj ◦ π, (3.16) for j ∈ {1,...,n}. Now, using (3.15) and (3.16) in (2.10) we have shown that for any projectable function f = f ◦ π, the Hamiltonian semimartingale Γh satisﬁes that n n 1 f ◦ π Γh − f ◦ π Γh = g grad f, Y π ◦ Γh dBj + Hess f (Y , Y ) π ◦ Γh dt, (3.17) 0 j s 2 j j j=1 Z j=1 Z X X or equivalently n 1 f ◦ π Γh − f ◦ π Γh − ∆ (f) π ◦ Γh dt = g grad f, Y π ◦ Γh dBj . (3.18) 0 2 M j s Z j=1 Z X L´azaro and Ortega: Stochastic Hamiltonian dynamical systems 19

n h i h Since i=1 g grad f, Yj (Γ )dB is a local martingale (see [P90, Theorem 20, page 63]), π(Γ ) is a Brownian motion. P R Brownian motions on Lie groups. Let now G be a (finite dimensional) Lie group with Lie algebra g and assume that G admits a bi-invariant metric g, for example when G is Abelian or compact. This metric induces a pairing in g invariant with respect to the adjoint representation of G on g. Let {ξ1,...,ξn} be an orthonormal basis of g with respect to this invariant pairing and let {ν1,...,νn} be ∗ the corresponding dual basis of g . The infinitesimal generator vector fields {ξ1G,...,ξnG} defined by ξiG(h) = TeLh · ξ, with Lh : G → G the left translation map, h ∈ G, i ∈ {1,...n}, are obviously an 1 orthonormal parallelization of G, that is g(ξiG, ξjG) := δij . Since g is bi-invariant then ∇X Y = 2 [X, Y ], for any X, Y ∈ X(G) (see [O83, Proposition 9, page 304]), and hence ∇ξiG ξiG = 0. Therefore, in this particular case the first component h0 of the Hamiltonian function introduced in (3.14) is zero and we 1 n can hence take hG = (h1, ..., hn) and XG = Bt , ..., Bt when we consider the Hamilton equations that define the Brownian motion with respect to g. As a special case of the previous construction that serves as a particularly simple illustration, we are going to explicitly build the Brownian motion on a circle. Let S1 = {eiθ | θ ∈ R} be the unit circle. The stochastic Hamiltonian differential equation for the semimartingale Γh associated to 1 1 X : R+ × Ω → R, given by Xt(ω) := Bt(ω), and the Hamiltonian function h : TS ≃ S × R → R iθ iθ given by h(e , λ) := λ, is simply obtained by writing (3.17) down for the functions f1(e ) := cos θ and iθ h h h f2(e ) := sin θ which provide us with the equations for the projections X and Y of Γ onto the OX and OY axes, respectively. A straightforward computation yields 1 1 dXh = −Y hdB − Xhdt and dY h = XhdB − Y hdt, (3.19) 2 2 which, incidentally, coincides with the equations proposed in expression (5.1.13) of [Ok03]. A solution h h h iBt of (3.19) is (Xt , Yt ) = (cos Bt, sin Bt), that is, Γt = e . Brownian motions on arbitrary manifolds. Let (M,g) be a not necessarily parallelizable Rie- mannian manifold. In this case we will reproduce the same strategy as in the previous paragraphs but replacing the cotangent bundle of the manifold by the cotangent bundle of its orthonormal frame bundle. Let Ox (M) be the set of orthonormal frames for the tangent space TxM. The orthonormal frame bundle O (M) = x∈M Ox (M) has a natural smooth manifold structure of dimension n (n + 1)/ 2. We denote by π : O (M) → M the canonical projection. We recall that a curve γ : (−ε,ε) ⊂ R → S O (M) is called horizontal if γt is the parallel transport of γ0 along the projection π (γt). The set of tangent vectors of horizontal curves that contain a point u ∈ O (M) defines the horizontal subspace HuO (M) ⊂ TuO (M) , with dimension n. The projection π : O (M) → M induces an isomorphism Tuπ : HuO (M) → Tπ(u)M. On the orthonormal frame bundle, we have n horizontal vector fields Yi, i = 1, ..., n, defined as follows. For each u ∈ O (M), let Yi (u) be the unique horizontal vector in HuO (M) such that Tuπ (Yi) = ui, where ui is the ith unit vector of the orthonormal frame u. Now, given a smooth function F ∈ C∞ (O (M)), the operator

∆O(M) (F )= Yi [Yi [F ]] i=1 X is called Bochner’s horizontal Laplacian on O (M). At the same time, we recall that the Laplacian ∞ ∆M (f), for any f ∈ C (M), is deﬁned as ∆M (f) = Tr(Hess f). These two Laplacians are related by the relation ∗ ∆O(M) (π f) = ∆M (f) , (3.20) for any f ∈ C∞ (M) (see [H02]). L´azaro and Ortega: Stochastic Hamiltonian dynamical systems 20

The Eells-Elworthy-Malliavin construction of Brownian motion can be summarized as follows. Con- sider the following stochastic diﬀerential equation on O (M) (see [IW89]):

n i δUt = Yi (Ut) δBt (3.21) i=1 X where Bj , j = 1, ..., n, are n-independent Brownian motions. Using the conventions introduced in the appendix 6.4 the expression (3.21) is the Stratonovich stochastic diﬀerential equation associated to the Stratonovich operator:

n e (v,u) : TvR −→ TuO (M) n i n i v = i=1 v ei 7−→ i=1 v Yi (u) ,

n where {e1,...,en} is a fixed basis for R .P A solution of theP stochastic differential equation (3.21) is called a horizontal Brownian motion on O (M) since, by the Itôformula,

n n 1 F (U) − F (U )= Y [F ] (U ) δBi = Y [F ] (U ) dBi + ∆ (F ) (U ) ds, 0 i s s i s s 2 O(M) s i=1 i=1 X Z X Z Z for any F ∈ C∞ (O (M)). In particular, if F = π∗ (f) for some f ∈ C∞ (M) , by (3.20)

n 1 f (X) − f (X )= Y [π∗ (f)] (U ) dBi + ∆ f (X ) ds, 0 i s s 2 M s i=1 X Z Z where Xt = π (Ut), which implies precisely that Xt is a Brownian motion on M. ∗ In order to generate (3.21) as a Hamilton equation, we introduce the functions hi : T O (M) → R, ∗ i =1, ..., n, given by hi (α)= hα, Yii. Recall that T O (M) being a cotangent bundle it has a canonical symplectic structure. Mimicking the computations carried out in the parallelizable case it can be seen that the Hamiltonian vector ﬁeld Xhi coincides with Yi when acting on functions of the form ∞ ∗ F ◦ πT ∗O(M), where F ∈ C (O (M)) and πT ∗O(M) is the canonical projection πT ∗O(M) : T O (M) → h O (M). By (2.8), the Hamiltonian semimartingale Γ associated to h = (h1, ..., hn) and to the stochastic ∗ 1 n Hamiltonian equations on T O(M) with stochastic component X = Bt , ..., Bt is such that

h h F ◦ πT ∗O(M) Γ − F ◦ πT ∗O(M) Γ0 n n h i h i = F ◦ πT ∗O(M),hi Γs δBs = Yi [F ] πT ∗O(M) Γs δBs i=1 Z i=1 Z X X ∞ h h for any F ∈ C (O (M)). This expression obviously implies that U = πT ∗O(M) Γ is a solution of (3.21) and consequently Xh = π U h is a Brownian motion on M. 3.5 The inverted pendulum with stochastically vibrating suspension point The equation of motion for small angles of a damped inverted unit mass pendulum of length l with a vertically vibrating suspension point is y¨ g φ¨ = + φ − λφ,˙ (3.22) l l where φ is the angle that measures the separation of the pendulum from the vertical upright position, y = y (t) is the height of the suspension point (externally controlled), λ is the friction coefficient, and g Lázaro and Ortega: Stochastic Hamiltonian dynamical systems 21 is the gravity constant. By construction, the point (φ, φ˙)=(0, 0) corresponds to the upright equilibrium position. It can be shown that if the function y(t) is of the form y(t) = az (ωt), with z periodic, the amplitude a is sufficiently small, and the frequency ω is sufficiently high, then this equilibrium becomes nonlinearly stable. We now consider the case in which the external forcing of the suspension point is given by a continuous 2 stochastic processz ˙ : R+ × Ω → R such thatz ˙ is continuous and stationary. Under this assumptions, the equation (3.22) becomes the stochastic differential equation g dφ = φdt,˙ dφ˙ = φ − λφ˙ dt + ε2ω2φdz˙ , (3.23) l t where ε := a/l. Observe that this equation is not Hamiltonian unless the friction term −λφ˙ vanishes (λ = 0), in which case one obtains a Hamiltonian stochastic system with Hamiltonian function h(φ, φ˙)= 1 2 ˙2 p2 1 2 2 2 1 2 ( 2 (l φ − lφ ), 4 (ε ω φl) , − 2 (εωφl) ) and noise semimartingale Xt = (t, [z, ˙ z˙], z˙) (the symplectic form is obviously l2dφ ∧ dφ˙). The stability of the upright position of the stochastically forced pendulum has been studied in [O06, I01], and references therein. In [O06] it is assumed that the noise has the fairly strong mixing property. We recall that a continuous, adapted, stationary process Γ : R+ × Ω → R has the fairly strong mixing 2 ∞ property if E Γt < ∞, there exists a real function c such that 0 c (s) ds < ∞, and for any t>s R kE [Γt − E [Γt]|Fs]kL2 ≤ c (t − s) kΓs − E [Γs]kL2 ,

2 where k·kL2 stands for the L norm. For example, if x is the unique stationary solution with zero mean of the Itôequations dxt = ytdt, dyt = − (xt + yt) dt + dBt, 2 1 2 1 where Bt is a standard Brownian motion, thenx ˙ t − 2 = yt − 2 has the fairly strong mixing property. Using this hypothesis, it can be shown [O06, Theorem 1] that if z : R+ × Ω → R is a continuously differentiable and stationary process such that, for any t ∈ R+, E [zt] = 0, E [exp (ε |zt|)] < ∞ if ε = a/l is sufficiently small, and the processz ˙2 has the fairly strong mixing property, then the solution (φ, φ˙)=(0, 0) of (3.23) is exponentially stable in probability, if ε is sufficiently small and g < E z˙2 . p lε4 Moreover, Ovseyevich shows in [O06, Section 4] that if we put λ = 0 in (3.23) and we consider hence the inverted pendulum as a Hamiltonian system, then the equilibrium point (φ, φ˙)=(0, 0) is unstable.

4 Critical action principles for the stochastic Hamilton equations

Our goal in this section is showing that the stochastic Hamilton equations can be characterized by a variational principle that generalizes the one used in the classical deterministic situation. In the following pages we shall consider an exact symplectic manifold (M,ω), that is, there exist a one-form θ ∈ Ω (M) such that ω = −dθ. The archetypical example of an exact symplectic manifold is the cotangent bundle T ∗Q of any manifold Q, with θ the Liouville one-form. In the following pages we will proceed in two stages. In the first subsection we will construct a critical action principle based on using variations of the solution semimartingale using the flow of a vector field on the manifold. Even though this approach is extremely natural and mathematically very tractable it yields a variational principle (Theorem 4.9) that does not fully characterize the stochastic Hamilton’s equations. In order to obtain such a characterization one needs to use more general variations associated to the flows of vector fields defined on the solution semimartingale, that is, they depend on Ω. This complicates considerably the formulation and will be treated separately in the second subsection. Lázaro and Ortega: Stochastic Hamiltonian dynamical systems 22

Deﬁnition 4.1 Let (M,ω = −dθ) be an exact symplectic manifold, X : R+ × Ω → V a semimartingale taking values on the vector space V , and h : M → V ∗ a Hamiltonian function. We denote by S (M) and S (R) the sets of M and real-valued semimartingales, respectively. We deﬁne the stochastic action associated to h as the map S : S(M) → S(R) given by

S (Γ) = hθ,δΓi− h (Γ) ,δX , Z Z D E ∗ where in the previous expression, h (Γ) : R+ × Ω → V × V b is given by h (Γ) (t,ω) := (Xt(ω),h(Γt(ω))).

4.1 Variations involvingb vector fields on the phase spaceb Definition 4.2 Let M be a manifold, F : S (M) → S (R) a map, and Γ ∈ S (M). A local one-parameter group of diffeomorphisms ϕ : D⊂ R × M → M is said to be complete with respect to Γ if there exists ǫ> 0 such that ϕs(Γ) is a well-defined process for any s ∈ (−ǫ,ǫ). We say that F is differentiable at Γ in the direction of a local one parameter group of diffeomorphisms ϕ complete with respect to Γ, if for any sequence {sn} N ⊂ R such that sn −→ 0, the family n∈ n→∞ 1 Xn = (F (ϕsn (Γ)) − F (Γ)) sn d converges uniformly on compacts in probability (ucp) to a process that we will denote by ds s=0 F (ϕs (Γ)) and that is referred to as the directional derivative of F at Γ in the direction of ϕ . s

Remark 4.3 Note that global one-parameter groups of diffeomorphisms (for instance, flows of complete vector fields) are complete with respect to any semimartingale. Let Γ : R+ × Ω → M be a M-valued continuous and adapted stochastic process and A ⊂ M a set. We will denote by τA = inf {t> 0 | Γt (ω) ∈/ A} the first exit time of Γ with respect to A. We recall that τA is a stopping time if A is a Borel set. Additionally, let Γ be a semimartingale and K a compact set such that Γ0 ⊂ K. Then, any local one-parameter group of diffeomorphisms ϕ is complete with respect to the stopped process ΓτK . Note that this conclusion could also hold for certain non-compact sets. The proof of the following proposition can be found in Section 5.1. Proposition 4.4 Let M be a manifold, α ∈ Ω (M) a one-form, and F : S (M) → S (R) the map defined by F (Γ) := hα, δΓi. Then F is differentiable in all directions. Moreover, if Γ : R+ × Ω → M is a continuous semimartingale, ϕ is an arbitrary local one-parameter group of diffeomorphisms complete with respect toR Γ, and Y ∈ X(M) is the vector field associated to ϕ, then d d d F (ϕ (Γ)) = hα, δ (ϕ ◦ Γ)i = hϕ∗α, δΓi = h£ α, δΓi . (4.1) ds s ds s ds s Y s=0 s=0 Z s=0 Z Z

The symbol £Y α denotes the Lie derivative of α in the direction given by Y . Corollary 4.5 In the setup of Definition 4.1 let α = ω♭ (Y ) ∈ Ω(M), with ω♭ the inverse of the vector ♯ ∗ bundle isomorphism ω : T M → TM induced by ω. Let Γ : R+ × Ω → M be a continuous adapted semimartingale. ϕ an arbitrary local one-parameter group of diffeomorphisms complete with respect to Γ, and Y ∈ X(M) the associated vector field. Then, the action S is differentiable at Γ in the direction of ϕ and the directional derivative is given by d S (ϕ (Γ)) = − hα, δΓi− dh ω# (α) (Γ) ,δX + i θ (Γ) − i θ (Γ ) . (4.2) ds s Y Y 0 s=0 Z Z

L´azaro and Ortega: Stochastic Hamiltonian dynamical systems 23

Proof. It is clear from Proposition 4.4 that

1 s hϕ∗θ − θ,δΓi −→→0 h£ θ,δΓi s s Y Z Z in ucp. The proof of that result can be easily adapted to show that ucp

1 s ϕ∗h − h (Γ) ,δX −→→0 £ h (Γ) ,δX . s s Y Z D E Z D E Thus, using (6.5) and α = ω♭ (Y ) ∈bΩ(Mb ), b

d S (ϕ (Γ)) = h£ θ,δΓi− £ h (Γ) ,δX = hi dθ + d (i θ) ,δΓi− hdh (Y ) (Γ) ,δXi ds s Y Y Y Y s=0 Z Z Z Z D E b # = − hα, δΓi + hd (iY θ) ,δΓi− dh ω (α) (Γ) ,δX Z Z Z # = − hα, δΓi− dh ω (α) (Γ) ,δX + (iY θ) (Γ) − (iY θ)(Γ0) . Z Z Corollary 4.6 (Noether’s theorem) In the setup of Definition 4.1, let ϕ : R × M → M be a one parameter group of diffeomorphisms and Y ∈ X(M) the associated vector field. If the action S : S (M) → S (R) is invariant by ϕ, that is, S (ϕs (Γ)) = S (Γ), for any s ∈ R, then the function iY θ is a conserved quantity of the stochastic Hamiltonian system associated to h : M → V ∗.

h Proof. Let Γ be the Hamiltonian semimartingale associated to h with initial condition Γ0. Since ϕs leaves invariant the action we have that d S ϕ Γh =0 ds s s=0 and hence by (4.2) we have that

h # h h 0= − α, δΓ − dh ω (α) Γ ,δX + iY θ Γ − iY θ (Γ0) . Z Z As Γh is the Hamiltonian semimartingale associated to h we have that

− α, δΓh = dh ω# (α) Γh ,δX Z Z h and hence iY θ Γ = iY θ (Γ0), as required.

Remark 4.7 The hypotheses of the previous corollary can be modiﬁed by requiring, instead of the ∞ invariance of the action by ϕs, the existence of a function F ∈ C (M) such that

d S ϕ Γh = F (Γ) − F (Γ ). ds s 0 s=0

In that situation, the conserved quantity is iY θ + F . Before we state the Critical Action Principle for the stochastic Hamilton equations we need one more deﬁnition. L´azaro and Ortega: Stochastic Hamiltonian dynamical systems 24

Definition 4.8 Let M be a manifold and A a set. We will say that a local one parameter group of diffeomorphisms ϕ : D × M → M fixes A if ϕs (y) = y for any y ∈ A and any s ∈ R such that X d (s,y) ∈ D. The corresponding vector field Y ∈ (M) given by Y (m) = ds s=0 ϕs(m) satisfies that Y | =0. A

Theorem 4.9 (First Critical Action Principle) Let (M,ω = −dθ) be an exact symplectic manifold, X : R+ × Ω → V a semimartingale taking values on the vector space V such that X0 = 0, and ∗ h : M → V a Hamiltonian function. Let m0 ∈ M be a point in M and Γ : R+ × Ω → M a continuous semimartingale such that Γ0 = m0. Let K be a compact set that contains the point m0. If the semimartingale Γ satisfies the stochastic Hamilton equations (2.7) (with initial condition Γ0 = m0) up to time τK then for any local one-parameter group of diffeomorphisms ϕ that fixes the set {m0} ∪ ∂K we have d 1 S (ϕ (ΓτK )) =0 a.s.. (4.3) {τK <∞} ds s s=0 τK

Proof. We start by emphasizing that when we write that Γ satisﬁes the stochastic Hamiltonian equations (2.7) up to time τK we mean that

τK hβ,δΓi + dh ω# (β) (Γ) ,δX =0. Z Z For the sake of simplicity in our notation we deﬁne the linear operator Ham : Ω(M) → S(R) given by

Ham (β) := hβ,δΓi + dh ω# (β) (Γ) ,δX , β ∈ Ω (M) . Z Z Suppose now that the semimartingale Γ satisfies the stochastic Hamilton equations up to time τK . Let ϕ be a local one-parameter group of diffeomorphisms that fixes {m0} ∪ ∂K, and let Y ∈ X(M) be the associated vector field. Then, taking α = ω♭ (Y ), we have by Corollary 4.5,

d S (ϕ (ΓτK )) = − hα, δΓτK i− dh ω# (α) (ΓτK ) ,δX + i θ (ΓτK ) , (4.4) ds s Y s=0 Z Z since Y (m0) = 0 and hence iY θ (Γ0) = 0. Additionally, since Γ is continuous, 1{τK <∞}ΓτK ∈ ∂K and Y |∂K = 0. Hence,

d 1 S (ϕ (ΓτK )) = −1 hα, δΓτK i + dh ω# (α) (ΓτK ) ,δX . {τK <∞} ds s {τK <∞} s=0 τK Z Z τK

Now, Proposition 5.5 and the hypothesis on Γ satisfying Hamilton’s equation guarantee that the previous expression equals

d τK 1 S (ϕ (ΓτK )) = −1 hα, δΓτK i + dh ω# (α) (ΓτK ) ,δX {τK <∞} ds s {τK <∞} s=0 τK Z Z τK τK # = −1{τK <∞} hα, δΓi + dh ω (α) (Γ) ,δX Z Z τK = −1{τK <∞} [Ham(α)τK ]=0 a.s., as required. L´azaro and Ortega: Stochastic Hamiltonian dynamical systems 25

Remark 4.10 The relation between the Critical Action Principle stated in Theorem 4.9 and the classical one for Hamiltonian mechanics is not straightforward since the categories in which both are formulated are very much different; more specifically, the differentiability hypothesis imposed on the solutions of the deterministic principle is not a reasonable assumption in the stochastic context and this has se- rious consequences. For example, unlike the situation encountered in classical mechanics, Theorem 4.9 does not admit a converse within the set of hypotheses in which it is formulated. In order to elaborate a little bit more on this question let (M,ω = −dθ) be an exact symplectic manifold, take the Hamiltonian function h ∈ C∞(M), and consider the stochastic Hamilton equations with trivial stochastic component X : R+ × Ω → R given by Xt(ω)= t. As we saw in Remark 2.7 the paths of the semimartingales that solve these stochastic Hamilton equations are the smooth curves that integrate the standard Hamilton equations. In this situation the action reads

S (Γ) = hθ,δΓi− h (Γs) ds. Z Z If the path Γt(ω) is differentiable then the integral hθ,δΓi (ω) reduces to the Riemann integral θ and S(Γ)(ω) coincides with the classical action. In particular, if Γ is a solution of the stochastic Γt(ω) R Hamilton equations then the paths Γ (ω) are necessarily differentiable (see Remark 2.7), they satisfy the R t standard Hamilton equations, and hence make the action critical. The following elementary example shows that the converse is not necessarily true, that is one may have semimartingales that satisfy (4.3) and that do not solve the Hamilton equations up to time τK . We will consider a deterministic example. Let m0, m1 ∈ M be two points. Suppose there exists an integral curve γ : [t0,t1] → M of the Hamiltonian vector field Xh defined on some time interval [t0,t1] such that γ (t0) = m0 and γ (t1) = m1. Define the continuous and piecewise smooth curve σ : [0,t1] → M as follows: m if t ∈ [0,t ] σ (t)= 0 0 γ (t) if t ∈ [t ,t ] . 0 1 Let ϕ be a local one-parameter group of diffeomorphisms that fixes {m0,m1}. Then by (4.2) d t S (ϕ (σ)) = − α + hα, X i (σ(t)) dt + hθ(σ(t)), Y (σ(t))i − hθ(m ), Y (m )i, ds s h 0 0 s=0 t Zσ|[0,t] Z0

d ♭ where Y (m) = ds s=0 ϕs (m), for any m ∈ M and α = ω (Y ). Using that σ satisﬁes the Hamilton equations on [t ,t ] and α (m ) = 0, it is easy to see that 0 1 0

d S (ϕ (σ)) =0, ds s s=0 t1

that is, σ makes the action critical. However, it does not satisfy the Hamilton equations on the interval [0,t1] , because they do not hold on (0,t0). This shows that the converse of the statement in Theorem 4.9 is not necessarily true. In the following subsection we will obtain such a converse by generalizing the set of variations allowed in the variational principle.

4.2 Variations involving vector ﬁelds on the solution semimartingale We start by spelling out the variations that we will use in order to obtain a converse to Theorem 4.9.

Definition 4.11 Let M be a manifold and Γ a M-valued semimartingale. Let s0 > 0; we say that the 0 map Σ : (−s0,s0) × R+ × Ω → M is a pathwise variation of Γ whenever Σt =Γt for any t ∈ R+ a.s.. We say that the pathwise variation Σ of Γ converges uniformly to Γ whenever the following properties are satisfied: Lázaro and Ortega: Stochastic Hamiltonian dynamical systems 26

(i) For any f ∈ C∞ (M), f (Σs) → f (Γ) in ucp as s → 0.

∞ (ii) There exists a process Y : R+ × Ω → TM over Γ such that, for any f ∈ C (M), the Stratonovich integral Y [f] δX exists for any continuous real semimartingale X (this is for instance guaranteed if Y is a semimartingale) and, additionally, the increments (f (Σs) − f (Γ))/ s converge in ucp to R Y [f] as s → 0. We will call such a Y the inﬁnitesimal generator of Σ. We will say that Σ (respectively, Y ) is bounded when its image lies in a compact set of M (respectively, TM). The next proposition shows that, roughly speaking, there exist bounded pathwise variations that converge uniformly to a given semimartingale with prescribed bounded inﬁnitesimal generator.

Proposition 4.12 Let Γ be a continuous M-valued semimartingale Γ, K ⊆ M a compact set, and τK τK the ﬁrst exit time of Γ from K. Let Y : R+ ×Ω → TM be a bounded process over Γ (that is, the image of Y lies in a compact subset of TM) such that Y [f] δX exists for any continuous real semimartingale X and for any f ∈ C∞ (M). Then, there exists a bounded pathwise variation Σ that converges uniformly R to ΓτK whose inﬁnitesimal generator is Y .

Proof. Let {(Vk, ϕk)}k∈N be a countable open covering of M by coordinate patches such that any Vk is contained in a compact set. This covering is always available by the second countability of the manifold and Lindelöf’s Lemma. Let {Uk}k∈N be an open subcovering such that, if Uk ⊆ Vi for some k, i ∈ N, then Uk ( Vi. Let {τm}m∈N be a sequence of stopping times (available by Lemma 3.5 in [E89]) such that, a.s., τ0 =0, τm ≤ τm+1, supm τm = ∞, and that, on each of the sets [τm, τm+1]∩{τm+1 > τm} the semimartingale Γ takes values in the open set Uk(m), for some k (m) ∈ N. Since K is compact, it can be covered by a finite number of these open sets, i.e. K ⊆ ∪j∈J Ukj , where |J| < ∞. 1 n Let xkj ≡ (xkj ,...,xkj ), n = dim (M) be a set of coordinate functions on Ukj (m) and (xkj , vkj ) ≡ 1 n 1 n −1 (xkj ,...,xkj , vkj ,...,vkj ) the corresponding adapted coordinates for TM on πT M Ukj (m) . Since Y is τK bounded and covers Γ , and on [τm, τm+1] ∩{τm+1 > τm} the semimartingale Γ takes values in the 1 open set Ukj (m), there exist a skj > 0 such that, on [τm, τm+1] ∩{τm+1 > τm}, the points (xkj (Γ) + 1 n n svkj (Y ) ,...,xkj (Γ) + svkj (Y )) lie in the image of some coordinate patch Vkj containing Ukj (m) for all s ∈ −skj ,skj . Let s0 = minj∈J skj . Now, since the sets of the form Im := [τm, τm+1) ∩ {τm > τm} ⊂ R × Ω, m ∈ N form a disjoint partition of R × Ω we define Σ as the map that for +1 + + any m ∈ N satisfies

Σ|Im : (−s0,s0) × [τm, τm+1) ∩{τm+1 > τm} −→ Vkj −1 (s,t,ω) 7−→ ϕk xkj (Γt (ω)) + svkj (Yt (ω)) . Observe that by construction the image of Σ is covered by a ﬁnite number of coordinated patches and therefore, by hypothesis, contained in a compact set. Σ is hence bounded. More speciﬁcally

s {Σt (ω) | (s,t,ω) ∈ (−s0,s0) × R × Ω}⊆ Vkj . (4.5) j J [∈ It is immediate to see that Σ is a pathwise variation which converges uniformly to ΓτK . Indeed, if ∞ f ∈ C (M) has compact support within one of the elements in the family Ukj j∈J , it can be easily checked that f (Σs) − f (Γ) f (Σs) −→ f (Γ) and −→ Y [f] . (4.6) ucp s ucp s→0 s→0 ∞ If, more generally, f ∈ C (M) has not compact support contained in one of the Ukj j∈J , observe that, by (4.5), we only need to consider the restriction of f to V . Take now a partition of the j∈J kj S L´azaro and Ortega: Stochastic Hamiltonian dynamical systems 27

unity {φk}k∈N subordinated to the covering {Uk}k∈N. Since {supp (φk)}k∈N is a locally ﬁnite family and j∈J Vkj is contained in a compact set because, by hypothesis, so is each Vkj for any j ∈ J, then among all the {φ } only a ﬁnite number of them have their supports in U , say {φ } with S k k∈N kj j∈J ki i∈I |I| < ∞. Thus, |I| f| = φ f ∪j∈J Vkj ki i=1 X and since each φki f is a function similar to those considered in (4.6) it is straightforward to see that those implications also hold for f. The following result generalizes Proposition 4.4 to pathwise variations of a semimartingale. The proof can be found in Section 5.2

Proposition 4.13 Let Γ be a M-valued continuous semimartingale Γ, K ⊆ M a compact set, and τK the ﬁrst exit time of Γ from K. Let Σ be a bounded pathwise variation that converges uniformly to ΓτK and Y : R+ × Ω → TM the inﬁnitesimal generator of Σ that we will also assume to be bounded. Then, for any α ∈ Ω (M),

1 s τK τK τK τK lim hα, δΣ i− hα, δΓ i = hiY dα, δΓ i + hα (Γ ) , Y i − hα (Γ ) , Y i . ucp s t=0 s→0 Z Z Z The next theorem shows that the generalization of the Critical Action Principle in Theorem 4.9 to pathwise variations fully characterizes the stochastic Hamilton’s equations. Theorem 4.14 (Second Critical Action Principle) Let (M,ω = −dθ) be an exact symplectic man- ∗ ifold, X : R+ × Ω → V a semimartingale that takes values in the vector space V, and h : M → V a Hamiltonian function. Let m0 be a point in M and Γ : R+ × Ω → M a continuous adapted semimartingale defined on [0, ζΓ) such that Γ0 = m0. Let K ⊆ M be a compact set that contains m0 and τK the first exit time of Γ from K. Suppose that τK < ∞ a.s.. Then, (i) For any bounded pathwise variation Σ with bounded infinitesimal generator Y which converges uniformly to ΓτK uniformly, the action has a directional derivative that equals d 1 S (Σs) := lim [S (Σs) − S (ΓτK )] ds s→0 s s=0

τK [ τK τK τK = hiY dθ,δΓ i− Y [h](Γ ),δX + hθ (Γ ) , Y i − hθ (Γ ) , Y it=0 , (4.7) Z Z D E where the symbol Y[[h](ΓτK ) is consistent with the notation introduced in Definition 4.1 (ii) The semimartingale Γ satisfies the stochastic Hamiltonian equations (2.7) with initial condition Γ0 = m0 up to time τK if and only if, for any bounded pathwise variation Σ : (−s0,s0)×R+ ×Ω → τK s M with bounded infinitesimal generator which converges uniformly to Γ and such that Σ0 = m0 s and ΣτK =ΓτK a.s. for any s ∈ (−s0,s0), d S (Σs) =0 a.s.. ds s=0 τK

Proof. We first show that the limit (4.7) exist. Let Σ be an arbitrary bounded pathwise variation converging to Γ uniformly and Y : R+ × Ω → TM its infinitesimal generator, that we also assume to be bounded. We have 1 1 1 [S (Σs) − S (ΓτK )] = hθ,δΣsi− hθ,δΓτK i − h (Σs) − h (ΓτK ) ,δX . (4.8) s s s Z Z Z D E b b Lázaro and Ortega: Stochastic Hamiltonian dynamical systems 28

By Proposition 4.13, the ﬁrst summand in the right hand side of (4.8) converges ucp to

τK τK τK hiY dθ,δΓ i + hθ (Γ ) , Y i − hθ (Γ ) , Y it=0 . Z as s → 0. An argument similar to the one leading to Proposition 4.13 shows that the second summand converges to Y[[h](ΓτK ),δX . Hence, D E 1 R s τK τK [ τK τK τK lim [S (Σ ) − S (Γ )] = hiY dθ,δΓ i− Y [h](Γ ),δX + hθ (Γ ) , Y i − hθ (Γ ) , Y i . s→0 s t=0 Z Z D E τK τK If we denote by η := −iY dθ = iY ω the one-form over Γ built using the vector ﬁeld Y over Γ , the previous relation may be rewritten as d S (Σs) = − hη,δΓτK i− dh (ΓτK ) ω# (η) ,δX +hθ (ΓτK ) , Y i−hθ (ΓτK ) , Y i . (4.9) ds t=0 s=0 Z Z

We are now going to prove the assertion in part (ii). Recall that the hypothesis that Γ satisﬁes the stochastic Hamilton equations up to time τK means that

τK hβ,δΓi + dh · ω# (β) (Γ) ,δX =0, (4.10) Z Z for any β ∈ Ω (M). We now show that this expression is also true if we replace β with any process ∗ η : R+ × Ω → T M over Γ such that the two Stratonovich integrals involved in (4.10) are well-deﬁned (for instance if β is a semimartingale). Indeed, invoking ([E89, 7.7]) and Whitney’s embedding theorem, there exist an integer p ∈ N such that the manifold M can be seen as an embedded submanifold of Rp. In this embedded picture, there exists a family of functions f 1, ..., f p ⊂ C∞ (Rp) such that the one-form η may be written as p j η = Zjdf , j=1 X where the Zj : R+ × Ω → R, j ∈ {1, ..., p}, are real processes. Moreover, using the properties of the Stratonovich integral (see [E89, Proposition 7.4]),

τK hη,δΓi + dh · ω# (η) (Γ) ,δX

Z Z τK p j # j = Zj δ df ,δΓ + dh · ω df (Γ) ,δX   j=1 Z Z Z X p τK j # j = Zjδ df ,δΓ + dh · ω df (Γ) ,δX , j=1 Z Z Z X where the last equality follows from Proposition 5.5. Therefore, since df j is a deterministic one-form τ we can conclude that df j ,δΓ + dh · ω# df j (Γ) ,δX K = 0, which justiﬁes why (4.10) also holds if we replace β ∈ Ω (M) by an arbitrary integrable one-form η over Γ. R R Suppose now that Γ satisﬁes the stochastic Hamilton equations up to τK and let Σ : (−s0,s0) × R+ × Ω → M be a pathwise variation like in the statement of the theorem. We want to show that d S (Σs) = 0 a.s.. ds s=0 τK

L´azaro and Ortega: Stochastic Hamiltonian dynamical systems 29

Due to (4.9), we have that d S (Σs) = − hη,δΓτK i + dh (ΓτK ) ω# (η) ,δX +hθ (ΓτK ) , Y i −hθ (ΓτK ) , Y i . ds τK t=0 s=0 τK Z Z τK s s Since Σ0 = m0 and ΣτK = ΓτK a.s. for any s ∈ (−s0,s0), then Y0 = YτK = 0 a.s. and both hθ (ΓτK ) , Y i and hθ (ΓτK ) , Y i vanish. Moreover, τK t=0

τK hη,δΓτK i + dh (ΓτK ) ω# (η) ,δX = hη,δΓτK i + dh (ΓτK ) ω# (η) ,δX Z Z τK Z Z τK τK = hη,δΓi + dh (Γ) ω# (η) ,δX Z Z τK (4.11) which is zero because of (4.10). In the last equality we have used Proposition 5.5. Conversely, suppose that d S (Σs) = 0 a.s. for arbitrary bounded pathwise variations ds s=0 τK τK tending to Γ uniformly, like in the statement. We want to show that (4.10) holds. Since our pathwise variations satisfy that Y0 = YτK = 0 a.s., we obtain that d S (Σs) = − hη,δΓτK i + dh (ΓτK ) ω# (η) ,δX = 0 (4.12) ds s=0 τK Z Z τK where η is an arbitrary bounded one form over Γ. Suppose now that η is a semimartingale. Then ∗ 1[0,t]η : R+ × Ω → T M is again bounded and expressions

τK τK # 1[0,t]η,δΓ and dh (Γ ) ω 1[0,t]η ,δX Z Z are well-deﬁned by Proposition 5.7 because both ΓτK and X are continuos semimartingales. We already saw in (4.11) that (4.12) is equivalent to

hη,δΓi + dh (Γ) ω# (η) ,δX =0. Z Z τK Replacing η by 1[0,t]η in (4.12) and using again the Proposition 5.7, we write

t # # 0= 1[0,t]η,δΓ + dh (Γ) ω 1[0,t]η ,δX = hη,δΓi + dh (Γ) ω (η) ,δX τK ! Z Z Z Z τK τK = hη,δΓi + dh (Γ) ω# (η) ,δX = hη,δΓi + dh (Γ) ω# (η) ,δX . Z Z t∧τK Z Z t τ Since t is arbitrary this implies that the process hη,δΓi + dh (Γ) ω# (η) ,δX K is identically zero, as required. R R 5 Proofs and auxiliary results 5.1 Proof of Proposition 4.4 Before proving the proposition, we recall a technical lemma dealing with the convergence of sequences in a metric space. L´azaro and Ortega: Stochastic Hamiltonian dynamical systems 30

Lemma 5.1 Let (E, d) be a metric space. Let {xn}n∈N be a sequence of functions xn : (0,δ) → E where (0,δ) ⊂ R is an open interval of the real line. Suppose that xn converges uniformly on (0,δ) to ∗ a function x. Additionally, suppose that for any n, the limits lim xn (s) = xn ∈ E exist and so does s→0 ∗ lim xn. Then n→∞ ∗ lim x (s) = lim xn. s→0 n→∞ Proof. Let ε> 0 be an arbitrary real number. We have

∗ ∗ ∗ ∗ d x (s) , lim xn ≤ d (x (s) , xk (s)) + d (xk (s) , xk)+ d xk, lim xn . n→∞ n→∞ From the deﬁnition of limit and since xk (s) converges uniformly to x on (0,δ) , we can choose k0 such ∗ ∗ ε ε that d (xk, limn→∞ xn) < 3 and d (x (s) , xk (s)) < 3 , simultaneously for any k ≥ k0. Additionally, since ∗ ∗ ε lims→0 xk (s)= xk we choose s0 small enough such that d (xk (s) , xk) < 3 , for any s

∗ d x (s) , lim xn <ε n→∞ ∗ H for any s 0 is arbitrary, we conclude that lim x (s) = lim xn. s→0 n→∞ Proof of Proposition 4.4. First of all, the second equality in (4.1) is a straightforward consequence of [E89, page 93]. Now, let {Uk}k∈N be a countable open covering of M by coordinate patches. By [E89, Lemma 3.5] there exists a sequence {τm}m∈N of stopping times such that τ0 =0, τm ≤ τm+1, supm τm = ∞, a.s., and that, on each of the sets

[τm, τm+1] ∩{τm < τm+1} := {(t,ω) ∈ R+ × Ω | τm+1 (ω) > τm (ω) and t ∈ [τm (ω) , τm+1 (ω)]} (5.1) the semimartingale Γ takes its values in one of the elements of the family {Uk}k∈N. Second, the statement of the proposition is formulated in terms of Stratonovich integrals. However, the proof will be carried out in the context of Itˆointegration since we will use several times the notion of uniform convergence on compacts in probability (ucp) which behaves well only with respect to this integral. Regarding this point we recall that by the very deﬁnition of the Stratonovich integral of a 1-form α along a semimartingale Γ we have that

∗ ∗ hϕs α, δΓi = hd2 (ϕsα) , dΓi and h£Y α, δΓi = hd2 (£Y α) , dΓi . (5.2) Z Z Z Z The proof of the proposition follows directly from Lemma 5.1 by applying it to the sequence of functions given by 1 τn x (s) := [d (ϕ∗α) − d (α)] , dΓ . n s 2 s 2 Z This sequence lies in the space D of càglàd processes endowed with the topology of the ucp convergence. We recall that this space is metric [P90, page 57] and hence we are in the conditions of Lemma 5.1. In the following points we verify that the rest of the hypotheses of this result are satisfied.

(i) The sequence of functions {xn(s)}n∈N converges uniformly to

1 x(s) := [d (ϕ∗α) − d (α)] , dΓ . s 2 s 2 Z The pointwise convergence is a consequence of part (i) in Proposition 5.6. Moreover, in the proof of that result we saw that if d : D × D → R+ is a distance function function associated to the ucp convergence, L´azaro and Ortega: Stochastic Hamiltonian dynamical systems 31

then for any t ∈ R+ and any s ∈ (0,ǫ), d(xn(s), x(s)) ≤ P ({τn < t}). Since the right hand side of this inequality does not depend on s and P ({τn

By the construction of the covering {Uk}k∈N and of the stopping times {τm}m∈N, there exists a k(m) ∈ N such that the semimartingale Γ takes its values in Uk(m) when evaluated in the stochastic interval 1 ∗ s→0 (τn, τn+1] ⊂ [τn, τn+1] ∩{τn < τn+1}. Now, since d2 is a linear operator and s ((ϕs α) − α) (m) −→ 1 ∗ s→0 £Y α(m), for any m ∈ M, we have that s (d2 (ϕsα) − d2α) (m) −→ d2 (£Y α) (m). Moreover, a straight- 1 ∗ s→0 forward application of Taylor’s theorem shows that s (d2 (ϕs α) − d2α) |Uk(m) −→ d2 (£Y α) |Uk(m) uni- ∗ formly, using a Euclidean norm in τ Uk(m) (we recall that Uk(m) is a coordinate patch). This fact 1 ∗ s→0 immediately implies that 1(τn,τn+1] s (d2 (ϕsα) − d2α) (Γ) −→ 1(τn,τn+1]d2 (£Y α) (Γ) in ucp. As by construction the Itˆointegral behaves well when we apply it to a ucp convergent sequence of processes we have that

1 ∗ lim 1(τn,τn ] (d2 (ϕsα) − d2α) (Γ), dΓ = 1(τn,τn ] hd2 (£Y α) (Γ), dΓi . (5.3) ucp +1 s +1 s→0 Z Z Consequently,

τn 1 ∗ lim [d2 (ϕsα) − d2 (α)] , dΓ ucp s s→0 Z n−1 τm+1 τm 1 ∗ 1 ∗ = lim [d2 (ϕsα) − d2 (α)] , dΓ − [d2 (ϕs α) − d2 (α)] , dΓ ucp s s s→0 m=0 X Z Z n−1 n−1 1 ∗ = lim 1(τm,τm+1] (d2 (ϕs α) − d2α) , dΓ = 1(τm,τm+1] hd2 (£Y α) , dΓi ucp s s→0 m=0 Z m=0 Z X τn X = hd2 (£Y α) , dΓi , Z where in the second equality we have used Proposition 5.5 and the third one follows from (5.3). (iii) ∗ lim xn = hd2 (£Y α) , dΓi . n→∞ Z It is a straightforward consequence of Proposition 5.6. ∗ The equation (4.1) follows from Lemma 5.1 applied to the sequences {xn}n∈N and {xn}n∈N, and using the statements in (i), (ii), and (iii).

5.2 Proof of Proposition 4.13 We will start the proof by introducing three preparatory results.

Lemma 5.2 Let {Xn}n∈N and {Yn}n∈N be two sequences of real valued processes converging in ucp to a couple of processes X and Y respectively. Suppose that, for any t ∈ R+, the random variables supn∈N sup0≤s≤t |(Xn)s| and sup0≤s≤t |Ys| are bounded (their images lie in a compact set of R). Then, the sequence XnYn converges in ucp to XY as n → ∞. L´azaro and Ortega: Stochastic Hamiltonian dynamical systems 32

Proof. We need to prove that for any ε> 0 and any t ∈ R+,

P sup |(XnYn) − (XY ) |≤ ε −→ 1. s s n→∞ 0≤s≤t First of all, note that

sup |(XnYn)s − (XY )s|≤ sup |Xn| |Yn − Y | + sup |Y | |Xn − X| . 0≤s≤t 0≤s≤t 0≤s≤t

Hence, we have

sup |(XnYn)s − (XY )s|≤ ε ⊇ sup |Xn| |Yn − Y | + sup |Y | |Xn − X|≤ ε 0≤s≤t 0≤s≤t 0≤s≤t ε ε ⊇ sup |X | |Y − Y |≤ ∩ sup |Y | |X − X|≤ . n n 2 n 2 0≤s≤t 0≤s≤t Denote ε ε A := sup |X | |Y − Y |≤ , and B := sup |Y | |X − X|≤ , n n n 2 n n 2 0≤s≤t 0≤s≤t and let c be a constant such that supn∈N sup0≤s≤t |(Xn)s| < c and sup0≤s≤t |Ys| < c, available by the boundedness hypothesis. Then, ε 1 ≥ P (An) ≥ P sup |Yn − Y |≤ −→ 1, 2c n→∞ 0≤s≤t ε 1 ≥ P (Bn) ≥ P sup |Xn − X|≤ −→ 1. 2c n→∞ 0≤s≤t

Thus, P (An) → 1 and P (Bn) → 1 as n → ∞. But as P (An ∩ Bn)= P (An)+ P (Bn) − P (An ∪ Bn), we conclude that P (An ∩ Bn) −→ 1. n→∞

Since An ∩ Bn ⊆ sup0≤s≤t |(XnYn)s − (XY )s|≤ ε , we obtain P sup |(XnYn) − (XY ) |≤ ε −→ 1. H s s n→∞ 0≤s≤t

Lemma 5.3 Let {Xn}n∈N be a sequence of real processes converging in ucp to a process X. Let τ be a stopping time such that τ < ∞ a.s.. Then, the sequence of random variables {(Xn)τ }n∈N converge in probability to (X)τ . Proof. First of all we show that since τ < ∞ a.s., then P ({τ >t}) converges to zero as t → ∞. By contradiction, suppose that this is not the case. Then, denoting An := {τ>n}, we have that An+1 ⊆ An, so P (An) forms a non-increasing sequence of real numbers in the interval [0, 1]. Since this sequence is bounded below, it must have a limit. This limit corresponds to the probability of the event {τ = ∞}. If it is strictly positive then there is a contradiction with the fact that τ < ∞ a.s.. So P ({τ >t}) tends to zero as t → ∞. We now prove the statement of the lemma. Take some ε > 0 and an auxiliary t ∈ R+. The set {|(Xn)τ − Xτ | >ε} can be decomposed as the disjoint union of the following two events,

({|(Xn)τ − Xτ | >ε}∩{τ ≤ t}) ({|(Xn)τ − Xτ | >ε}∩{τ>t}) . [ L´azaro and Ortega: Stochastic Hamiltonian dynamical systems 33

The ﬁrst one is contained in the set sup0≤s≤t |(Xn)s − Xs| >ε whose probability, by hypothesis, converges to zero as n → ∞. Regarding the second one,

P ({|(Xn)τ − Xτ | >ε}∩{τ >t}) ≤ P ({τ >t}) . But P ({τ >t}) can be made arbitrarily small by taking the auxiliary t big enough. In conclusion, for any ε> 0, P ({|(Xn) − Xτ | >ε}) −→ 0 τ n→∞ in probability. H

Lemma 5.4 Let {Xn}n∈N be a sequence of real processes converging in ucp to a real process X and τ τ τ a stopping time. Then, the stopped sequence {Xn}n∈N converges in ucp to X as well.

Proof. We just need to observe that, for any t ∈ R+,

τ τ sup |(Xn)s − Xs|≤ ε ⊆ sup |(Xn)s − Xs |≤ ε . 0≤s≤t 0≤s≤t

Hence, since by hypothesis P sup0≤s≤t |(Xn)s − Xs|≤ ε converges to 1 as n → ∞, then so does P sup |(Xτ ) − Xτ |≤ ε . H 0≤s≤t n s s We now proceed with the proof of the proposition. We will start by using Whitney’s Embedding Theorem and the remarks in [E89, §7.7] to visualize M as an embedded submanifold of Rp , for some p ∈ N, and to write down our Stratonovich integrals as real Stratonovich integrals. Indeed, there exists a family of functions h1, ..., hp ⊂ C∞ (Rp) such that, in the embedded picture, the one form α can be p j ∞ p written as α = j=1 Zj dh , where Zj ∈ C (R ) for j ∈ {1, ..., p}. Therefore, using the properties of the Stratonovich integral (see [E89, Proposition 7.4]), P p 1 1 hα, δΣsi− hα, δΓτK i = Z (Σs) δ hj (Σs) − Z (ΓτK ) δ hj (ΓτK ) . (5.4) s s j j Z Z j=1 Z Z X p s j τK Adding and subtracting the term j=1 Zj (Σ ) δh (Γ ) in the right hand side of (5.4), we have

P R p 1 1 hα, δΣsi− hα, δΓτK i = Z (Σs) δhj (Σs) − Z (Σs) δhj (ΓτK ) s s j j j=1 Z Z X Z Z (1) p |1 {z } + (Z (Σs) − Z (ΓτK )) δhj (ΓτK ) . (5.5) s j j j=1 X Z (2)

We are going to study the terms (1) and (2) separately.| We star{zt by considering }

n n n σn = 0= T0 ≤ T1 ≤ . . . ≤ Tkn < ∞ , a sequence of random partitions that tends to the identity (in the sense of [P90, page 64]). L´azaro and Ortega: Stochastic Hamiltonian dynamical systems 34

1 s j s s j τK The expression (1): We want to study the ucp convergence of s Zj (Σ ) δh (Σ ) − Zj (Σ ) δh (Γ ) as s → 0. Deﬁne R R kn−1 1 1 n n s s j s Ti+1 j s Ti xn (s) := Zj (Σ )T n + Zj (Σ )T n h (Σ ) − h (Σ ) s 2 i+1 i i=0 X kn−1 n n 1 s s j τK Ti+1 j τK Ti − Zj (Σ )T n + Zj (Σ )T n h (Γ ) − h (Γ ) 2 i+1 i i=0 ! X k n n n n n−1 j s Ti+1 j τK Ti+1 j s Ti j τK Ti 1 s s h (Σ ) − h (Γ ) h (Σ ) − h (Γ ) = Zj (Σ )T n + Zj (Σ )T n − . 2 i+1 i s s i=0 ! X

1 s j s s j τK which corresponds to the discretization of the Stratonovich integrals s Zj (Σ ) δh (Σ )− Zj (Σ ) δh (Γ ) using the random partitions of σn. Indeed, by [P90, Corollary 1, page 291], R R

1 s j s s j τK xn (s) −→ Zj (Σ ) δh (Σ ) − Zj (Σ ) δh (Γ ) . ucp s n→∞ Z Z

n On the other hand, as Ti < ∞ a.s. for any i ∈ {1, ..., kn}, part (i) in Deﬁnition 4.11 and Lemma 5.3 imply that

1 s s 1 τK τK Zj (Σ )T n + Zj (Σ )T n −→ Zj (Γ )T n + Zj (Γ )T n 2 i+1 i ucp 2 i+1 i s→0 The convergence above is in probability but, for convenience, we prefer to regard these random variables as trivial processes. Furthermore, part (ii) in Deﬁnition 4.11 and Lemma 5.4

n n n j s T j τ T j s j τ Ti+1 h (Σ ) i+1 − h (Γ K ) i+1 h (Σ ) − h (Γ K ) T n = −→ Y hj i+1 , s s ucp s→0 n n n j s T j τ T j s j τ Ti h (Σ ) i − h (Γ K ) i h (Σ ) − h (Γ K ) T n = −→ Y hj i s s ucp s→0 by Deﬁnition (4.11) item 3 and Lemma (5.4). Now, since by hypothesis Σ and Y are bounded then so 1 s s j j j are Zj (Σ ) n + Zj (Σ ) n and Y h = iY dh (dh is only evaluated on the compact K since Y 2 Ti+1 Ti τK is a vector ﬁeld over Γ ) and hence by Lemma 5.2

kn−1 T n T n 1 τK τK j i+1 j i ∗ xn (s) −→ Zj (Γ )T n + Zj (Γ )T n Y h − Y h =: xn ucp 2 i+1 i i=0 s→0 X In addition, by [P90, Corollary 1, page 291],

∗ τK j xn −→ Zj (Γ ) δ Y h . ucp n Z →∞ Hence, by Lemma 5.1 we conclude that

1 s j s s j τK τK j Zj (Σ ) δh (Σ ) − Zj (Σ ) δh (Γ ) −→ Zj (Γ ) δ Y h . (5.6) s ucp Z Z Z s→0 L´azaro and Ortega: Stochastic Hamiltonian dynamical systems 35

1 s τK j τK The expression (2): We want to study now the ucp convergence of s (Zj (Σ ) − Zj (Γ )) δh (Γ ) as s → 0. As in the previous paragraphs, we deﬁne R kn−1 1 1 n n s s j τK Ti+1 j τK Ti yn (s) := Zj (Σ )T n + Zj (Σ )T n h (Γ ) − h (Γ ) s 2 i+1 i i=0 X kn−1 n n 1 τK τK j τK Ti+1 j τK Ti − Zj (Γ )T n + Zj (Γ )T n h (Γ ) − h (Γ ) 2 i+1 i i=0 ! X kn−1 s τK s τK Zj (Σ ) n − Zj (Γ ) n Zj (Σ ) n − Zj (Γ ) n n n 1 Ti+1 Ti+1 T T T T = + i i hj (ΓτK ) i+1 − hj (ΓτK ) i 2 s s i=0 ! X 1 s τK j τK as a discretization of the Stratonovich integral s (Zj (Σ ) − Zj (Γ )) δh (Γ ) using σn. Then, by construction, R 1 s τK j τK yn (s) −→ (Zj (Σ ) − Zj (Γ )) δh (Γ ) . ucp s n→∞ Z On the other hand, invoking Deﬁnition 4.11 and Lemma 5.3 we have that

s τK Zj (Σ ) n − Zj (Γ ) n s τK Ti+1 Ti+1 Zj (Σ ) − Zj (Γ ) = −→ Y [Zj ] n ucp Ti+1 s s T n i+1 s→0 s τK Zj (Σ ) n − Zj (Γ ) n s τK Ti Ti Zj (Σ ) − Zj (Γ ) = −→ Y [Zj ] n . ucp Ti s s T n i s→0

We now use again the boundedness of Σ and Y to guarantee the boundedness of Y [Zj ] n = (iY dZj ) n Ti+1 Ti+1 and Y [Zj] n = (iY dZj) n (notice that dZj is only evaluated on the compact set K because Y is a Ti Ti vector ﬁeld over ΓτK ⊆ K). Therefore, by Lemma 5.2,

kn−1 n n 1 j τK Ti+1 j τK Ti ∗ xn (s) −→ Y [Zj ]T n + Y [Zj]T n h (Γ ) − h (Γ ) := xn. ucp 2 i+1 i i=0 s→0 X

∗ j τK Additionally, the sequence {xn}n∈N obviously converge in ucp to Y [Zj ] δ h (Γ ) as n → ∞. Hence, by Lemma 5.1, we conclude that R

1 s τK j τK j τK (Zj (Σ ) − Zj (Γ )) δh (Γ ) −→ Y [Zj ] δ h (Γ ) . (5.7) s ucp Z Z s→0 To sum up, if we substitute (5.6) and (5.7) in (5.5) we obtain that

p 1 s τK τK j j τK hα, δΣ i− hα, δΓ i −→ Zj (Γ ) δ Y h + Y [Zj] δ h (Γ ) . s ucp Z Z j=1 Z Z s→0 X Using the integration by parts formula,

τK j τK j τK j j τK Zj (Γ ) δ Y h = Zj (Γ ) Y h − Zj (Γ ) Y h t=0 − Y h δ (Zj (Γ )) Z Z τK τK j τK = hα (Γ ) , Y i − hα (Γ ) , Y it=0 − Y h δ (Zj (Γ )) Z L´azaro and Ortega: Stochastic Hamiltonian dynamical systems 36 and, consequently,

1 s τK j τK j τK hα, δΣ i− hα, δΓ i −→ Y [Zj ] δ h (Γ ) − Y h δ (Zj (Γ )) s ucp s→0 R R τK R τK R + hα (Γ ) , Y i − hα (Γ ) , Y it=0 . In order to conclude the proof, we claim that

j τK j τK τK Y [Zj] δ h (Γ ) − Y h δ (Zj (Γ )) = hiY dα, δΓ i . (5.8) Z Z Z Indeed,

p p p j j j j dα = d Zj dh = dZj ∧ dh , and iY dα = Y [Zj] dh − Y h dZj   j=1 j=1 j=1 X X X   which proofs (5.8), as required.

5.3 Auxiliary results about integrals and stopping times In the following paragraphs we collect three results that are used in the paper in relation with the interplay between stopping times and integration limits.

Proposition 5.5 Let X be a continuous semimartingale deﬁned on [0, ζX ) and Γ a continuous semimartingale. Let τ, ξ be two stopping times such that τ ≤ ξ < ζX . Then,

τ τ ξ τ (X · Γ) = 1[0,τ]X · Γ = (X · Γ ) and (X · Γ) − (X · Γ) = 1(τ,ξ]X · Γ

An equivalent result holds when dealing with the Stratonovich integral, namely

τ τ XδΓ = XδΓτ = Xτ δΓ . Z Z Z τ τ Proof. By [P90, Theorem 12, page 60] we have that 1[0,τ]X · Γ = (X · Γ) = (X · Γ ). Therefore,

ξ τ (X · Γ) − (X · Γ) = 1[0,ξ]X · Γ − 1[0,τ]X · Γ= 1[0,ξ] − 1[0,τ] X · Γ= 1(τ,ξ]X · Γ.

As to the Stratonovich integral, since X and Γ are semimartingales, we can write [P90, Theorem 23, page 68] that

τ 1 1 XδΓ = (X · Γ)τ + [X, Γ]τ = (X · Γτ )+ [X, Γτ ]= XδΓτ . 2 2 Z Z Finally, observe that for any process, (Xτ )τ = Xτ . On the other hand, taking into account that τ 1[0,τ]X = 1[0,τ]X and [Γ,X] = [X, Γ], we have

τ 1 1 τ XδΓ = 1 X · Γ+ [X, Γ]τ = 1 Xτ · Γ+ [X, Γ]τ [0,τ] 2 [0,τ] 2 Z 1 τ 1 τ τ = (Xτ · Γ)τ + [Xτ , Γ] = Xτ · Γ+ [Xτ , Γ] = Xτ δΓ . 2 2 Z L´azaro and Ortega: Stochastic Hamiltonian dynamical systems 37

Proposition 5.6 Let X : R+ × Ω → R be a real valued process. Let {τn}n∈N be a sequence of stopping times such that a.s. τ0 =0, τn ≤ τn+1, for all n ∈ N, and supn∈N τn = ∞. Then, X = lim Xτn . ucp n→∞

In particular, if Γ : R+ × Ω → M is a continuous M-valued semimartingale and η ∈ Ω2 (M) then,

τk k−1 hη, dΓi = lim hη, dΓi = lim 1(τ ,τ ] hη, dΓi . ucp ucp n n+1 n=0 Z k→∞ Z k→∞ X Z

Proof. Let ε> 0 and t ∈ R+. Then for any s ∈ [0,t] one has

τn {|X − X|s >ε}⊆{τn

Hence for any t ∈ R+ τn P ({|X − X|s >ε}) ≤ P ({τn

The result follows because P ({τn

τk k−1 τn+1 τn k−1

hη, dΓi = hη, dΓi − hη, dΓi = 1(τn,τn+1] hη, dΓi n=0 n=0 Z X Z Z X Z and the result follows.

Proposition 5.7 Let X and Y be two real semimartingales. Suppose that X is continuous and X0 =0. t Then, for any t ∈ R+, the Stratonovich integral 1[0,t]Y δX is well defined and equal to YδX . R 1 R Proof. If 1[0,t]Y δX was well defined, it should be equal to 1[0,t]Y dX + 2 1[0,t]Y,X . Since 1 Y dX is well defined, the only thing that we need to check is that 1 Y,X exists. On the [0,t] R R [0,t] other hand, recall that ([P90, Theorem 12 page 60 and Theorem 23 page 68]) R t 1 1 YδX = 1 Y dX + [Y,X]t = 1 Y dX + Y t,X . [0,t] 2 [0,t] 2 Z Z Z t Hence, what we are actually going to proceed by showing that 1[0,t]Y,X is equal to [Y ,X]. n n n Let σn = 0= T ≤ T ≤ ... ≤ T < ∞ be a sequence of random partitions tending to the identity 0 1 kn (in the sense of [P90, page 64]). Given two real processes X and Y , their quadratic variation, if it exists, can be defined as the limit in ucp when n → ∞ of the following sums

kn−1 T n T n T n T n [Y,X] = lim Y i+1 − Y i X i+1 − X i . ucp i=0 n→∞ X Let now

kn−1 n n n n t Ti+1 t Ti T T Hn := Y − Y X i+1 − X i , i=0 X kn−1 T n T n n n i+1 t i Ti Ti Gn := 1[0,t]Y − 1[0,t]Y X +1 − X . i=0 X L´azaro and Ortega: Stochastic Hamiltonian dynamical systems 38

t It is clear that the sequence {Hn}n∈N converges uniformly on compacts in probability to [Y ,X]. We are going to prove that there exists such a convergence for the sequence of processes {Gn}n∈N by showing that the elements (Gn)s coincide with (Hn)s, for any s ∈ R+, up to a set whose probability tends to zero as n → ∞. We will consider two cases: n n 1. The case s ≤ t. Given a speciﬁc i ∈ {0, ..., kn − 1}, and recalling that by construction Ti ≤ Ti+1 n n t Ti+1 t Ti a.s., it is clear that (Y ) − (Y ) = YT n ∧s − YT n∧s is diﬀerent from 0 only for those ω ∈ Ω in s i+1 i n {Ti

n n YTi+1∧s − YTi . (5.9)

n n Ti+1 t Ti n On the other hand, 1[0,t]Y − 1[0,t]Y is again diﬀerent from 0 only in the set {Ti t. In this case, (Y ) − (Y ) = Yt∧T n − Yt∧T n which is diﬀerent from 0 s i+1 i n only in the set {Ti

n n Yt∧Ti+1 − YTi (5.10)

n n Ti+1 t Ti However, in this case 1 Y − 1 Y = 1 n Y n −1 n Y n , which is equal [0,t] [0,t] T ≤t t∧Ti+1 T ≤t t∧Ti s { i+1 } { i } n n n n to (5.10) in the set Ti+1 ≤ t (which contains {Ti

n n Ti+1 t Ti where it takes the value −YT n . For any other ω ∈ Ω not in these sets, 1[0,t]Y − 1[0,t]Y (ω)= i s n 0. Therefore, whenever s>t, (Gn)s and (Hn)s are diﬀerent only for the ω ∈ Ai (t). Observe that, since t is ﬁxed, only one of the sets {An (t)} is non-empty and, on it, i i∈{0,...,kn−1}

n (Hn)s − (Gn)s = Yt Xt − XTi . To sum up, the analysis that we just carried out shows that for any u ∈ R+

sup |(H ) − (G ) | = 1 n |Y | X − X n n s n s Ai (t) t t Ti 0≤s≤u for some i ∈{0, ..., kn − 1}. If X is continuous, this expression tells us that sup0≤s≤u |(Hn)s − (Gn)s| → 0 a.s. as n → ∞ which, in turn, implies that sup0≤s≤u |(Hn)s − (Gn)s| converges to 0 in probability as well. That is, for any ε> 0,

P sup |(Hn)s − (Gn)s| >ε → 0, as n → ∞, 0≤s≤u which is the same as saying that Hn − Gn converges to 0 in ucp. Thus, since Gn = Hn − (Hn − Gn) and the limit in ucp as n → ∞ exist for the both sequences {Hn}n∈N and {Hn − Gn}n∈N, so does the limit of {Gn}n∈N which, by deﬁnition, is the quadratic variation 1[0,t]Y,X . Moreover, as (Hn − Gn) → 0 in ucp as n → ∞, t Y ,X = lim Hn = lim Gn = 1[0,t]Y,X , ucp ucp n→∞ n→∞ which concludes the proof. L´azaro and Ortega: Stochastic Hamiltonian dynamical systems 39

6 Appendices 6.1 Preliminaries on semimartingales and integration In the following paragraphs we state a few standard definitions and results on manifold valued semimartingales and integration. Semimartingales are the natural setup for stochastic differential equations and, in particular, for the equations that we handle in this paper. For proofs and additional details the reader is encouraged to check, for instance, with [CW90, Du96, E89, IW89, LeG97, P90], and references therein. Semimartingales. The first element in our setup for stochastic processes is a probability space (Ω, F, P ) together with a filtration {Ft | t ≥ 0} of F such that F0 contains all the negligible events (complete filtration) and the map t 7−→ Ft is right-continuous, that is, Ft = ǫ>0 Ft+ǫ. A real-valued martingale Γ : R+ = [0, ∞) × Ω → R is a stochastic process such that for every pair T t,s ∈ R+ such that s ≤ t, we have:

(i) Γ is Ft-adapted, that is, Γt is Ft-measurable.

(ii) Γs = E[Γt |Fs].

(iii) Γt is integrable: E[|Γt|] < +∞.

p p For any p ∈ [1, ∞), Γ is called a L -martingale whenever Γ is a martingale and Γt ∈ L (Ω), for p p each t. If supt∈R+ E[|Γt| ] < ∞, we say that Γ is L -bounded. The process Γ is locally bounded if for any time t ≥ 0, sup{|Γs(ω)| | s ≤ t} < ∞, almost surely. Every continuous process is locally bounded. Recall that a process is said to be continuous when its paths are continuous. Most processes considered in this paper will be of this kind. Given two continuous processes X and Y we will write X = Y when they are a modification of each other or when they are indistinguishable since these two concepts coincide for continuous processes. A random variable τ : Ω → [0, +∞] is called a stopping time with respect to the filtration {Ft | t ≥ 0} if for every t ≥ 0 the set {ω | τ(ω) ≤ t} belongs to Ft. Given a stopping time τ, we define

Fτ = {Λ ∈ F | Λ ∩{τ ≤ t}∈Ft for any t ∈ R+} .

Given an adapted process Γ, it can be shown that Γτ is Fτ -measurable. Furthermore, the stopped process Γτ is defined as τ Γt := Γt∧τ := Γt1{t≤τ} +Γτ 1{t>τ}. A continuous local martingale is a continuous adapted process Γ such that for any n ∈ N, τn Γ 1{τn>0} is a martingale, where τn is the stopping time τn := inf{t ≥ 0 | |Γt| = n}. We say that the stochastic process Γ : R+ × Ω → R has finite variation whenever it is adapted and has bounded variation on compact subintervals of R+. This means that for each fixed ω ∈ Ω, the path t 7−→ Γt(ω) has bounded variation on compact subintervals of R+, that is, the supremum p sup i=1 |Γti (ω) − Γti−1 (ω)| over all the partitions 0 = t0 < t1 < · · · < tp = t of the interval [0,t] is finite. APcontinuous semimartingale is the sum of a continuous local martingale and a process with finite variation. It can be proved that a given semimartingale has a unique decomposition of the form Γ=Γ0 + V +Λ, with Γ0 the initial value of Γ, V a finite variation process, and Λ a local continuous semimartingale. Both V and Λ are null at zero.

The Itôintegral with respect to a continuous semimartingale. Let Γ : R+ × Ω → R be a continuous local martingale. It can be shown that there exists a unique increasing process with finite 2 variation [Γ, Γ]t such that Γt − [Γ, Γ]t is a local continuous martingale. We will refer to [Γ, Γ]t as the Lázaro and Ortega: Stochastic Hamiltonian dynamical systems 40

′ ′ ′ ′ quadratic variation of Γ. Given Γ=Γ0 + V +Λ, Γ =Γ0 + V +Λ two continuous local martingales we deﬁne their joint quadratic variation or quadratic covariation as 1 [Γ, Γ′] = ([Λ + Λ′, Λ+Λ′] − [Λ, Λ] − [Λ′, Λ′] ) . t 2 t t t

Let {Xn}n∈N be a sequence of processes. We will say that {Xn}n∈N converges uniformly on compacts in probability (abbreviated ucp) to a process X if for any ε> 0 and any t ∈ R+,

P sup |Xn − X|s >ε −→ 0, 0≤s≤t as n → ∞. Following [P90], we denote by L the space of processes X : R+ × Ω → R whose paths are left- continuous and have right limits. These are usually called c`agl`ad processes, which are initials in French for left-continuous with right limits. We say that a process X ∈ L is elementary whenever it can be expressed as p−1

X = X01{0} + Xi1(τi,τi+1], i=1 X where 0 ≤ τ1 < · · · < τp−1 < τp are stopping times, and X0 and Xi are F0 and Fτi -measurable random variables, respectively such that |X0| < ∞ and |Xi| < ∞ a.s. for all i ∈ {1,...,p − 1}. 1(τi,τi+1] is the characteristic function of the set (τi, τi+1] = {(t,ω) ∈ R+ × Ω | t ∈ (τi (ω) , τi+1 (ω)]} and 1{0} of {(t,ω) ∈ R+ × Ω | t =0} . It can be shown (see [P90, Theorem 10, page 57]) that the set of elementary processes is dense in L in the ucp topology. Let Γ be a semimartingale such that Γ0 = 0 and X elementary. We deﬁne Itˆo’s stochastic integral of X with respect to Γ as given by

p−1 τi+1 τi X · Γ := XdΓ := Xi(Γ − Γ ). (6.1) i=1 Z X In the sequel we will exchangeably use the symbols X · Γ and XdΓ to denote the Itôstochastic integral. It is a deep result that, if Γ is a semimartingale, the Itôstochastic integral is a continuous map from L into the space of processes whose paths are right-continuouR s and have left limits (càdlàg), usually denoted by D, equipped also with the ucp topology. Therefore we can extend the Itôintegral to the whole L. In particular, we can integrate any continuous adapted processes with respect to any semimartingale. Given any stopping time τ we define

τ XdΓ := (X · Γ)τ . Z0 τ τ It can be shown that (1[0,τ]X) · Γ = (X · Γ) = X · Γ . If there exists a stopping times ζΓ such that the semimartingale Γ is defined only on the stochastic intervals [0, ζΓ), then we may define the Itôintegral τ of X with respect to Γ on any interval [0, τ] such that τ < ζΓ by means of X · Γ . The Stratonovich integral and stochastic calculus. Given Γ and X two semimartingales we define the Stratonovich integral of X along Γ as

t t 1 XδΓ= XdΓ+ [X, Γ] . 2 t Z0 Z0 L´azaro and Ortega: Stochastic Hamiltonian dynamical systems 41

Let X1,...,Xp be p continuous semimartingales and f ∈ C2(Rp). The celebrated Itˆoformula states that p t ∂f f(X1,...,Xp)= f(X1,...,Xp)+ (X1,...,Xp)dXi t t 0 0 ∂xi s s s i=1 Z0 p X 1 t ∂2f + (X1,...,Xp)d[Xi,Xj ] 2 ∂xi∂xj s s s i,j=1 0 X Z The analogue of this equality for the Stratonovich integral is

p t ∂f f(X1,...,Xp)= f(X1,...,Xp)+ (X1,...,Xp)δXi. t t 0 0 ∂xi s s s i=1 0 X Z An important particular case of these relations are the integration by parts formulas

t t 1 XdΓ = (XΓ)t − (XΓ)0 − ΓdX − [X, Γ]t, 0 0 2 Z t Z t XδΓ = (XΓ)t − (XΓ)0 − ΓδX. Z0 Z0

1 p Stochastic differential equations. Let Γ = (Γ ,..., Γ ) be p semimartingales with Γ0 = 0 and f : Rq × Rp → Rq a smooth function. A solution of the Itôstochastic differential equation

p i i j dX = fj (X, Γ)dΓ (6.2) j=1 X 1 q 1 q with initial condition the random vector X0 = (X0 ,...,X0 ) is a stochastic process Xt = (Xt ,...,Xt ) i i p t i j q such that Xt − X0 = j=1 0 fj (X, Γ)dΓ . It can be shown [P90, page 310] that for any x ∈ R there q exists a stopping time ζ : R × Ω → R+ and a time-continuous solution X(t,ω,x) of (6.2) with initial P R condition x and deﬁned in the time interval [0, ζ(x, ω)). Additionally, lim supt→ζ(x,ω) kXt(ω)k = ∞ a.s. on {ζ < ∞} and X is smooth on x in the open set {x | ζ(x, ω) > t}. Finally, the solution X is a semimartingale.

6.2 Second order vectors and forms In the paragraphs that follow we review the basic tools on second order geometry needed in the definition of the stochastic integral of a form along a manifold valued semimartingale. The reader interested in the proofs of the statements cited in this section is encouraged to check with [E89], and references therein. Let M be a finite dimensional, second-countable, locally compact Hausdorff (and hence paracompact) manifold. Given m ∈ M, a tangent vector at m of order two with no constant term is a differential operator L : C∞ (M) −→ R that satisfies

L f 3 (m)=3f (m) L f 2 (m) − 3f 2 (m) L [f] (m) .

The vector space of tangent vectors of order two at m is denoted as τmM. The manifold τM := m∈M τmM is referred to as the second order tangent bundle of M. Notice that the (first order) tangent bundle TM of M is contained in τM. A vector field of order two is a smooth section of S the bundle τM → M. We denote the set of vector fields order two by X2(M). If Y,Z ∈ X(M) then the Lázaro and Ortega: Stochastic Hamiltonian dynamical systems 42

product ZY ∈ X2(M). Conversely, every second order vector field L ∈ X2(M) can be written as a finite sum of fields of the form ZY and W , with Z,Y,W ∈ X(M). The forms of order two Ω2(M) are the smooth sections of the cotangent bundle of order two ∗ ∗ ∞ τ M := m∈M τmM. For any f,g,h ∈ C (M) and L ∈ X2(M) we define d2f ∈ Ω2(M) by d2f(L) := L[f], and d2f · d2g ∈ Ω2(M) as S 1 d f · d g[L] := (L [fg] − fL [g] − gL [f]) . 2 2 2 It is easy to show that for any Y,Z,W ∈ X(M), 1 d f · d g [ZY ]= (Z [f] Y [g]+ Z [g] Y [f]) and d f · d g [W ]=0. 2 2 2 2 2

∗ ∞ More generally, let αm,βm ∈ TmM and choose f,g ∈C (M) two functions such that df(m)= αm and dg(m) = βm. It is easy to check that (df · dg)(m) does not depend on the particular choice of f and g above and hence we can write αm · βm to denote (df · dg)(m). If α, β ∈ Ω(M) then we can define ∞ α · β ∈ Ω2(M) as (α · β) (m) := α(m) · β(m). This product is commutative and C (M)-bilinear. It can be shown that every second order form can be locally written as a finite sum of forms of the type df · dg and d2h. The d2 operator can also be defined on forms by using a result (Theorem 7.1 in [E89]) that claims that there exists a unique linear operator d2 : Ω(M) → Ω2(M) characterized by

d2 (df)= d2f and d2 (fα)= df · α + fd2α.

6.3 Stochastic integrals of forms along a semimartingale

Let M be a manifold. A continuous M-valued stochastic process X : R+×Ω → M is called a continuous M-valued semimartingale if for each smooth function f ∈ C∞(M), the real valued process f ◦ X is a (real-valued) continuous semimartingale. We say that X is locally bounded if the sets {Xs(ω) | 0 ≤ s ≤ t} are relatively compact in M for each t ∈ R+, a.s. ∗ Let X be a M-valued semimartingale and θ : R+ × Ω → τ M be a càglàd locally bounded process over X, that is, π ◦ θ = X, where π : τ ∗M → M is the canonical projection. It can be shown (see [E89, Theorem 6.24]) that there exists a unique linear map θ 7−→ hθ,dXi that associates to each such θ a continuous real valued semimartingale and that is fully characterized by the following properties: for any f ∈C∞ (M) and any locally bounded càglàd real-valuedR process K,

hd2f ◦ X,dXi = f (X) − f (X0) , and hKθ,dXi = Kd hθ,dXi . (6.3) Z Z Z Z

The stochastic process hθ,dXi will be called the Itôintegral of θ along X. If α ∈ Ω2(M), we will write in the sequel the Itôintegral of α along X, that is, hα ◦ X,dXi as hα, dXi. R The integral of a (0, 2)-tensor b on M along X is the image of the unique linear mapping b 7−→ b(dX, dX) onto the space of real continuous processes withR finite variationR that for all f,g, ∈ C∞(M) satisfies R (fb)(dX, dX)= (f ◦ X)d b(dX, dX) and (df ⊗ dg) (dX, dX) = [f ◦ X,g ◦ X]. (6.4) Z Z Z Z

If α ∈ Ω (M) and X is a semimartingale on M, the real semimartingale hd2α, dXi is called the Stratonovich integral of α along X and is denoted by hα,δXi. This definition can be generalized by R R Lázaro and Ortega: Stochastic Hamiltonian dynamical systems 43 taking β a T ∗M valued semimartingale over X and by defining the Stratonovich integral as the unique real valued semimartingale that satisfies the properties

hdf,δXi = f (X) − f (X0) , and hZβ,δXi = Z (X) δ hβ,δXi , (6.5) Z Z Z Z for any f ∈ C∞ (M) and any continuous real valued semimartingale Z. Finally, it can be shown that (see [E89, Proposition 6.31]) for any f,g ∈ C∞(M), 1 hdf · dg,dXi = [f (X) ,g (X)] . (6.6) 2 Z 6.4 Stochastic diﬀerential equations on manifolds The reader interested in the details of the material presented in this section is encouraged to check with the chapter 7 in [E89]. Let M and N be two manifolds. A Stratonovich operator from M to N is a family {e(x, y)}x∈M,y∈N such that e(x, y) : TxM → TyN is a linear mapping that depends smoothly on its two entries. Let ∗ ∗ ∗ e (x, y) : Ty N → Tx M be the adjoint of e(x, y). Let X be a M-valued semimartingale. We say that a N-valued semimartingale is a solution of the the Stratonovich stochastic diﬀerential equation

δY = e(X, Y )δX (6.7) if for any α ∈ Ω(N), the following equality between Stratonovich integrals holds:

hα, δY i = he∗(X, Y )α,δXi. Z Z It can be shown [E89, Theorem 7.21] that given a semimartingale X in M, a F0 measurable random variable Y0, and a Stratonovich operator e from M to N, there are a stopping time ζ > 0 and a solution Y of (6.7) with initial condition Y0 defined on the set {(t,ω) ∈ R+ × Ω | t ∈ [0, ζ(ω))} that has the following maximality and uniqueness property: if ζ′ is another stopping time such that ζ′ < ζ and Y ′ ′ ′ is another solution defined on {(t,ω) ∈ R+ × Ω | t ∈ [0, ζ (ω))}, then Y and Y coincide in this set. If ζ is finite then Y explodes at time ζ, that is, the path Yt with t ∈ [0, ζ) is not contained in any compact subset of N. The stochastic differential equations from the Itôintegration point of view require the notion of Schwartz operator whose construction we briefly review. The reader interested in the details of this construction is encouraged to check with [E89]. Note first that we can associate to any element L ∈ X2(M) a symmetric tensor L ∈ X(M) ⊗ X(M). Second, given x ∈ M and y ∈ N, a linear [ mapping from τxM into τyN is called a Schwartz morphism whenever f(TxM) ⊂ TyN and f(L)= b (f|TxM ⊗ f|TxM ) (L), for any L ∈ τxM. Third, let M and N be two manifolds; a Schwartz operator from M to N is a family {f(x, y)}x∈M,y∈N such that f(x, y) : τxM → τyN is a Schwartz operator that ∗ ∗ ∗ depends smoothlyb on its two entries. Let f (x, y) : τy N → τx M be the adjoint of f(x, y). Finally, let X be a M-valued semimartingale. We say that a N-valued semimartingale is a solution of the the Itô stochastic differential equation dY = f(X, Y )dX (6.8) if for any α ∈ Ω2(N), the following equality between Itôintegrals holds:

hα, dY i = hf ∗(X, Y )α, dXi. Z Z L´azaro and Ortega: Stochastic Hamiltonian dynamical systems 44

There exists an existence and uniqueness result for the solutions of these stochastic differential equations analogous to the one for Stratonovich differential equations. Given a Stratonovich operator e from M to N, there exists a unique Schwartz operator f : τM × N → τN defined as follows. Let γ(t) = (x(t),y(t)) ∈ M × N be a smooth curve that verifies e(x(t),y(t))(x ˙(t)) =y ˙(t), for all t. We define f(x(t),y(t)) Lx¨(t) := Ly¨(t) , where the second or- d2 der differential operators Lx¨(t) ∈ τx(t)M and Ly¨(t) ∈ τy(t)N are defined as Lx¨(t) [h] := dt2 h (x (t)) d2 ∞ ∞ and Ly¨(t) [g] := dt2 g (y (t)), for any h ∈ C (M) and g ∈ C (N). This relation completely deter- mines f since the vectors of the form Lx¨(t) span τx(t)M. Moreover, the Itôand Stratonovich equations δY = e(X, Y )δX and dY = f(X, Y )dX are equivalent, that is, they have the same solutions. Acknowledgments We thank Michel Emery and Jean-Claude Zambrini for carefully going through the paper and for their valuable comments and suggestions. We also thank Nawaf Bou-Rabee and Jerry Marsden for stimulating discussions on stochastic variational integrators. The authors acknowledge partial support from the French Agence National de la Recherche, contract number JC05-41465. J.-A. L.-C. acknowledges support from the Spanish Ministerio de Educación y Ciencia grant number BES- 2004-4914. He also acknowledges partial support from MEC grant BFM2006-10531 and Gobierno de Aragón grant DGA-grupos consolidados 225-206. J.-P. O. has been partially supported by a “Bonus QualitéRecherche” contract from the Universitéde Franche-Comté.

References

[AM78] Abraham, R., and Marsden, J.E. [1978] Foundations of Mechanics. Second edition, Addison-Wesley.

[A03] Arnold, L. [2003] Random Dynamical Systems. Springer Monographs in Mathematics. Springer Verlag.

[Ar89] Arnold, V.I. [1989] Mathematical Methods of Classical Mechanics. Second edition. Volume 60 of Graduate Texts in Mathematics, Springer Verlag.

[B81] Bismut, J.-M.[1981] M´ecanique Al´eatoire. Lecture Notes in Mathematics, volume 866. Springer-Verlag.

[BRO07] Bou-Rabee, N. and Owhadi, H. [2007] Stochastic Variational Integrators. arXiv:0708.2187.

[BRO07a] Bou-Rabee, N. and Owhadi, H. [2007] Stochastic Variational Partitioned Runge-Kutta Integrators for Con- strained Systems. arXiv:0709.2222.

[BJ76] Box, G. E. P. and Jenkins, G. M.[1976] Time Series Analysis: Forecasting and Control. Holden-Day.

[CH06] Chorin, A.J. and Hald, O.H.[2006] Stochastic Tools in Mathematics and Science. Surveys and Tutorials in the Applied Mathematical Sciences, volume 1. Springer Verlag.

[CW90] Chung, K. L. and Williams, R. J. [1990] Introduction to Stochastic Integration. Second edition. Probability and its Applications. Birkh¨auser Verlag.

[CD06] Cresson, J. and Darses, S. [2006] Plongement stochastique des syst`emes lagrangiens. C. R. Math. Acad. Sci. Paris, 342, 333-336.

[Du96] Durrett, R. [1996] Stochastic Calculus. A Practical Introduction. Probability and Stochastics Series. CRC Press.

[E89] Emery,´ M. [1989] Stochastic Calculus in Manifolds. Springer-Verlag.

[E90] Emery,´ M. [1990] On two transfer principles in stochastic differential geometry. Séminaire de Probabilités, XXIV, 1988/89, 407-441, Lecture Notes in Math., 1426, Springer Verlag. Correction: Séminaire de Probabilités, XXVI, 633, Lecture Notes in Math., 1526, Springer Verlag, 1992.

[G66] Gihman, I. I. [1966] Stability of solutions of stochastic diﬀerential equations. (Russian) Limit Theorems Statist. Inference (Russian) pp. 14–45 Izdat. ”Fan”, Tashkent. English translation: Selected Transl. Statist. and Prob- ability, 12, Amer. Math. Soc., pp. 125-154. MR 40, number 944. L´azaro and Ortega: Stochastic Hamiltonian dynamical systems 45

[Ha80] Hasminskii, R. Z. [1980] Stochastic Stability of Differential Equations. Translated from the Russian by D. Louvish. Monographs and Textbooks on Mechanics of Solids and Fluids: Mechanics and Analysis, 7. Sijthoff and Noordhoff, Alphen aan den Rijn—Germantown, Md.

[H02] Hsu, E. P. [2002] Stochastic Analysis on Manifolds. Graduate Studies in Mathematics, 38. American Mathe- matical Society.

[IW89] Ikeda, N. and Watanabe, S. [1989] Stochastic Diﬀerential Equations and Diﬀusion Processes. Second edition. North-Holland Mathematical Library, 24. North-Holland Publishing Co.

[I01] Imkeller, P. and Lederer, C. [2001] Some formulas for Lyapunov exponents and rotation numbers in two dimen- sions and the stability of the harmonic oscillator and the inverted pendulum. Dyn. Syst., 16(1), 29-61.

[K81] Kunita, H. [1981] Some extensions of Itˆo’s formula. Seminaire de probabilit´es de Strasbourg XV, 118-141. Lecture Notes in Mathematics, 850, Springer Verlag.

[LeG97] Le Gall, J.-F. [1997] Mouvement Brownian et Calcul Stochastique. Notes de Cours DEA 1996-97. Available at http://www.dma.ens.fr/ legall/.

[Ne67] Nelson, E. [1967] Dynamical Theories of Brownian Motion. Princeton University Press.

[LL76] Landau, L.D., and Lifshitz, E.M. [1976] Mechanics. Volume 1 of Course of Theoretical Physics. Third Edition. Pergamon Press.

[L06] Li, X.-M. An averaging principle for integrable stochastic Hamiltonian systems. Preprint available at http://www.lboro.ac.uk/departments/ma/research/preprints/papers06/06-27.pdf.

[M81] Meyer, P.-A. [1981] G´eom´etrie stochastique sans larmes. Seminar on Probability, XV (Univ. Strasbourg, Stras- bourg, 1979/1980), 44–102, Lecture Notes in Math., 850, Springer Verlag.

[M82] Meyer, P.-A. [1982] Géométrie différentielle stochastique. II. Seminar on Probability, XVI, Supplement, 165–207, Lecture Notes in Math., 921, Springer-Verlag.

[Ok03] Øksendal, B.[2003]Stochastic Diﬀerential Equations. Sixth Edition. Universitext. Springer-Verlag.

[O06] Ovseyevich, A. I. [2006] The stability of an inverted pendulum when there are rapid random oscillations of the suspension point. Journal of Applied Mathematics and Mechanics, 70, 762-768.

[O83] O’Neill, B.[1983] Semi-Riemannian Geometry. With Applications to Relativity Pure and Applied Mathematics, volume 103. Academic Press.

[LO07] L´azaro-Cam´ı, J.-A. and Ortega, J.-P. [2007] Reduction of stochastic Hamiltonian systems. In preparation.

[M99] Misawa, T. [1999] Conserved quantities and symmetries related to stochastic dynamical systems. Ann. Inst. Statist. Math., 51(4), 779–802.

[OP04] Ortega, J.-P. and Planas-Bielsa, V. [2004] Dynamics on Leibniz manifolds. J. Geom. Phys., 52(1), 1-27.

[P90] Protter, P. [2005] Stochastic Integration and Diﬀerential Equations. A New Approach. Applications of Mathe- matics, volume 21. Second Edition. Springer-Verlag.

[Sch82] Schwartz, L. [1982] Géométrie différentielle du 2ème ordre, semi-martingales et équations différentielles stochas- tiques sur une variétédifférentielle. Seminar on Probability, XVI, Supplement, 1–148, Lecture Notes in Math., 921, Springer Verlag.

[TZ97] Thieullen, M. and Zambrini, J. C. [1997] Probability and quantum symmetries. I. The theorem of Noether in Schrdinger’s Euclidean quantum mechanics. Ann. Inst. H. Poincar Phys. Thor., 67(3), 297–338.

[TZ97a] Thieullen, M. and Zambrini, J. C. [1997] Symmetries in the stochastic calculus of variations. Probab. Theory Related Fields, 107(3), 401–427.

[W80] Watanabe S. [1980] Differential and variation for flow of diffeomorphisms defined by stochastic differential equations on manifolds (in Japanese). S¯ukaiken K¯okyuroku, 391.

[Y81] Yasue, K. [1981] Stochastic calculus of variations. Journal of Functional Analysis, 41, 327-340. L´azaro and Ortega: Stochastic Hamiltonian dynamical systems 46

[ZY82] Zambrini, J.-C. and Yasue, K.[1982] Semi-classical quantum mechanics and stochastic calculus of variations. Annals of Physics, 143, 54-83.

[ZM84] Zheng, W. A. and Meyer, P.-A. [1984] Quelques r´esultats de ”m´ecanique stochastique”. Seminar on probability, XVIII, 223–244, Lecture Notes in Math., 1059, Springer-Verlag.